Proceedings of the SeaSAR 2023 Workshop, 2-6 May 2023, Svalbard, Norway

Organizing Committee Yves-Louis Desnos, ESA, Italy Catherine Downy, NERSC, Norway Marcus Engdahl, ESA, Italy Shridhar Jawak, SIOS, Norway Johnny A. Johannessen, NERSC, Norway Sabrina Lodadio, Serco c/o ESA, Italy Nicolas Longepe, ESA, Italy Jøran Idar Moen, UNIS, Norway Fabrizio Pera, Serco c/o ESA, Italy Lasse H. Pettersson, NERSC, Norway Diego Fernandez Prieto, ESA, Italy Eero Rinne, UNIS, Norway Ulla Vayrynen, Serco c/o ESA, Italy Scientific Committee Bertrand Chapron, Ifremer, France Fabrice Collard, OceanDataLab, France Maria Michela Corvino, ESA, Italy Yves-Louis Desnos, ESA, Italy Wolfgang Dierking, The Arctic University of Norway/Alfred Wegner Institute, Germany Craig Donlon, ESA, NL Alejandro Egido, ESA, NL Ralph Foster, University of Washington, Seattle, USA Lucile Gaultier, OceanDataLab, France Christine Gommenginger, NOC, UK Benjamin Holt, NASA JPL, USA Romain Husson, CLS, France Johnny A. Johannessen, NERSC, Norway Malin Johansson, The Arctic University of Norway, Norway Harald Johnsen, NORCE, Norway Cathleen Jones, NASA JPL, USA Marcel Kleinherenbrink, DELFT, NL Anton Korosov, NERSC, Norway Ronald Kwok, University of Washington Seattle, USA Paco Lopez-Dekker, Delft University of Technology, NL Artem Moiseev, NERSC, Norway Alexis Mouche, Ifremer, France Francisco Ocampo-Torres, CEMIE-Océano, Mexico William Perrie, Bedford Institute of Oceanography, CA Marcos Portabella, CSIC/ICM, Spain Diego Fernandez Prieto, ESA, Italy Roland Romeiser, RSMAS, University of Miami, USA Robert Shuchman, Michigan Technological Univ., USA Ad Stoffelen, KNMI, NL Justin Stopa, SOEST, USA Douglas Vandemark, Univ. of New Hampshire, USA Sponsored by: European Space Agency Publication: Proceedings of the SeaSAR 2023 Workshop, 2-6 May 2023, Svalbard, Norway Editors: Johnny A. Johannessen, Lasse Pettersson, Catherine Downy. Nansen Environmental Remote Sensing Center, Bergen, Norway DOI: 10.5281/zenodo.17294994 Contents Foreword List of Extended Abstracts: •Wave Retrievals •Near Surface Ocean Wind Retrievals and Detection of Extremes •Doppler Shift Retrievals •Sea Ice Retrievals •Sensor Synergy •Methodology and Techniques •Applications: Maritime Security •Future Missions List of Proceedings Papers: •Sea Ice Retrievals •Sensor Synergy •Methodology and Techniques •Applications List of Participants Foreword Twenty years after the workshop on Coastal and Marine applications of SAR the SeaSAR2023 workshop returned to Svalbard from 2-6 May 2023 (https://seasar2023.esa.int). The workshop was sponsored by ESA and organized jointly with the Nansen Environmental and Remote Sensing Center (NERSC), the University Centre of Svalbard and Svalbard Integrated Arctic Earth Observing System (SIOS). SeaSAR2023 was arranged around keynote presentations and thematic discussions in plenary followed by thematic panel working group discussion and reporting. The themes included: • Wave Retrievals • Near Surface Wind Retrievals and Detection of Extremes • Doppler Shift Retrievals • Sea Ice Retrievals • Sensor Synergy • Methodology and Techniques • Applications (Oil Spill, Ship Detection, etc) • Future Missions Covering these 8 themes the main workshop objectives included: • Review state-of-the-art in SAR-based geophysical parameter retrievals. • Identification of knowledge gaps and deficiencies. • Novel approaches for advancing scientific research and applications. • Importance and needs for validation. The workshop brought together around 80 experts and researchers (see list of participants) from around the world to share the latest advancements and challenges in coastal and marine applications of satellite-based Synthetic Aperture Radar (SAR) technology. In addition, 76 individuals attended the on-line WebEx transmissions of the plenary and the eight breakout sessions throughout the week. The YouTube broadcasting of the first day plenary session was viewed 148 times, mostly in real-time. Thanks to more than 3 decades of continuity of SAR missions the coastal and marine SAR scientific research and applications community have significantly advanced the quantitative understanding of SAR imaging capabilities. This is highlighted in these Proceedings of the SeaSAR 2023 regarding upper ocean currents and mesoscale structures, wave spectra, internal waves, sea ice field, surface slicks, near surface wind fields, marine atmosphere boundary layer processes and extremes as well as advances in methodologies and techniques and applications for oil spills and ship detection. Moreover, the outlook beyond 2030 and towards 2040 clearly evidences the continuity and sustainable access to data from future approved satellite SAR missions. However, as pointed out during the workshop there are deficiencies and challenges regarding design and implementation of a more comprehensive satellite SAR-based calibration-validation approach, jointly with systematic use of sensor synergy across the broad range of complementary satellite radar altimeters, scatterometry, interferometry, and highresolution optical radiometer and spectrometer. This will require dedicated international collaboration with involvement of the space agencies, in particular to enable multi-frequency and multi-look direction SAR-based observations with shorter imaging interval. Combined with continuous free and open available SAR data this will strengthen the generation of training datasets for machine learning development. List of Extended Abstracts Wave Retrievals Pleskachevsky, Andrey; Tings, Björn; Jacobsen, Sven Multiparametric Sea State Fields from Synthetic Aperture Radar using Method combining CWAVE Approach and Machine Learning Kleinherenbrink, Marcel; Ehlers, Frithjof; Gibert, Ferran; Hernandez, Sergi; Nouguier, Frederic; Chapron, Bertrand; Lopez-Dekker, Paco SAR altimetry cross-spectra Rikka, Sander; Nõmm, Sven; Alari, Victor; Björkqvist, Jan-Victor; Simon, Martin Wave Spectra, Spread, and Direction in the Baltic Sea from Sentinel-1 SAR Imagery Using the LSTM Model Khan, Salman Saeed Temporal Consistency of Sentinel-1 SAR Wave Mode Level-2 Ocean Swell Spectra Ocampo-Torres, Francisco J; Osuna, Pedro; Rascle, Nicolas G.; García-Nava, Héctor; Díaz Méndez, Guillermo M.; Esquivel-Trava, Bernardo; Villarreal-Olavarrieta, Carlos E.; MoraEscalante, Rodney; Hasimoto-Beltrán, Rogelio Assessment Of A Quasi-linear Inversion Scheme To Retrieve Ocean Surface Wave Spectrum from Synthetic Aperture Radar Image Spectrum: Results from Measurements in the Gulf of Mexico. Benchaabane, Amine; Husson, Romain; Mouche, Alexis; Frederic, Nouguier; Johnsen, Harald Wind-sea significant wave heights retrieval from Sentinel-1 with Deep Learning Nouguier, Frederic; Mouche, Alexis; Chapron, Bertrand; Kleinherenbrink, Marcel Ocean Wave Mapping By SAR: Numerical And Theoretical Perspectives Aouf, Lotfi; Hauser, Danièle; Collard, Fabrice; Chapron, Bertrand On The Complementary Assimilation Of SAR And SWIM Wave Spectra In The CMEMS Global Wave System Amir, Malik Muhammad Haris; Bogoni, Antonella; Dekker, Paco Lopez Unlocking the Potential of Distributed SwarmSAR for High-resolution Imaging of Ocean Waves Pouplin, Clément; Mouche, Alexis; Chapron, Bertrand; Yurovskaya, Maria; Filipot, JeanFrançois Tropical Cyclones Generated Waves: Processes And Contribution Of A Multi-Platform Approach Collard, Fabrice; Guitton, Gilles; López Radcenco, Manuel; Nouguier, Frederic; Chapron, Bertrand Swell Propagation and Forecast Across the North Atlantic Using Combined Sentinel1 Wave Mode, IW Mode and CFOSAT SWIM Observations. Near Surface Ocean Wind Retrievals & Detection of Extremes Stoffelen, Ad; Ni, Weicheng; Xu, Xingou; Portabella, Marcos; Makarova, Evgeniia; Cossu, Federico; Sánchez Rabaneda, Alberto Extreme Winds From SAR And Their Use For KuAnd C-Band Wind Scatterometer Resolution Enhancement Zecchetto, Stefano; Zanchetta, Andrea; Sclavo, Mauro; Shamsaddini, Parsa; Keshavarz, Ahmad High-Resolution SAR Winds from Deep Learning in Coastal Areas Khan, Salman; Young, Ian; Ribal, Agustinus; Canto, Marites; Davy, Robert; Hemer, Mark Australian Coastal SAR Ocean Winds: Data, Portal, and Next products Dimitriadou, Krystallia; Badger, Merete; Hasager, Charlotte Bay; Olsen, Bjarke Tobias SAR for Offshore Wind Fields in the Mediterranean Sea Hindberg, Heidi; Espeseth, Martine; Johnsen, Harald; Tollinger, Mathias Operational Wind Retrieval Using Cross-Polarization And Doppler Data Marquart, Robin; Mouche, Alexis; Chapron, Bertrand Analysis Of SAR Ocean Scenes Texture For Wind Direction Retrieval And Generation Of Synthetic Images Doppler Shift Retrievals Kleinherenbrink, Marcel; Yuan, Yan; Theodosiou, Andreas; Gaultier, Lucile; Collard, Fabrice; Chapron, Bertrand; Lopez-Dekker, Paco Multiscale Effects on Harmony's High-resolution Ocean Observations Elyouncha, Anis; Eriksson, Leif; Gommenginger, Christine Observations of the Agulhas Current by Along-track Interferometric Synthetic Aperture Radar Romeiser, Roland Review of TerraSAR-X Based Current Retrieval Activities at the University of Miami Martin, Adrien; Macedo, Karlus; McCann, David; Portabella, Marcos; Marié, Louis; Marquez, José; Carrasco, Ruben; Duarte, Rui; Meta, Adriano; Gommenginger, Christine; Martin-Iglesias, Petronilo; Casal, Tania OSCAR: A New Airborne Instrument To Image Ocean-Atmosphere Dynamics At The SubMesoscale Moiseev, Artem; Collard, Fabrice; Johannessen, Johnny A Ocean Surface Currents from Sentinel-1 Doppler observations Guitton, Gilles; Collard, Fabrice; Johnsen, Harald; Engen, Geir; Recchia, Andrea; Cotrufo, Alessandro; Bras, Sergio; Miranda, Nuno; Pinheiro, Muriel Towards Calibrated Sentinel-1 OCN RVL Products Domps, Baptiste; Guérin, Charles-Antoine Evaluation of Surface Currents Derived from Sentinel-1 SAR Doppler Shift in the Northwestern Mediterranean Sea Using Coastal HF Radars Sea Ice Retrievals Wiehle, Stefan; Murashkin, Dmitrii; Frost, Anja; König, Christine; König, Thomas Preliminary results of Sea Ice Classification using combined Sentinel-1 and Sentinel-3 data Karvonen, Juha Copernicus Marine Service SITAC SAR-Based Baltic Sea Ice Products Lohse, Johannes; Dierking, Wolfgang Combining CAnd L-band SAR Data For Automated Sea Ice Classification & Segmentation Frost, Anja; Imber, James; Murashkin, Dmitrii; Kortum, Karl; Gregorek, Daniel Towards Multitemporal Sea Ice Classification By Means Of Spaceborne SAR Image Time Series Wang, Qiang; Johansson, Malin; Lohse, Johannes; Doulgeris, Anthony P.; Eltoft, Torbjørn The Impact of Input Features in Deep Learning Based Sea Ice Mapping Eltoft, Torbjørn; Taelman, Catherine; Johansson, Malin; Lohse, Johannes; Gerland, Sebasitan; Dierking, Wolfgang The CIRFA-2022 Cruise to the Western Fram-Strait: Objectives, Ground Measurements, and Preliminary Results Li, Haiyan; Yang, Kun; Perrie, William Fine Sea Ice Classification with Gaofen-3 QuadPolarization SAR Observation Demchev, Denis; Eriksson, Leif E.B.; Hildeman, Anders; Dierking, Wolfgang Investigation of Multifrequency SAR Image Alignment by Ice Drift Compensation In The Marginal Ice Zone Johansson, Malin; Singha, Suman; Spreen, Gunnar; Howell, Stephen High resolution Land C-band polarimetric variability during MOSAiC Taelman, Catherine; Lohse, Johannes; Doulgeris, Anthony P. Tracking Backscatter Signatures Of Individual Sea Ice Floes Using In-Situ Ice Drift Observations Wulf, Tore; Buus-Hinkler, Jørgen; Singha, Suman; Kreiner, Matilde Brandt Operational SAR-based Sea Ice Concentration Retrieval Using Convolutional Neural Networks Korosov, Anton; Kleinherenbrink, Marcel Potential Application of the Earth Explorer 10 candidate Harmony for Sea Ice Model Validation Aparício, Sara A Multisensory SAR-Based Approach For Melt Ponds Retrievals Kim, Ekaterina; Skjetne, Roger; Høyland, Knut Vilhelm Quadruple Helix Framework for Sea Ice Monitoring: Next Steps Sensor Synergy Taillade, Thibault; Engdahl, Marcus; Fernandez, Diego Can We Retrieve Sea Surface Salinity with SAR Measurements? Zhang, Biao; Perrie, William; Zhang, Mingyu Polar Low Recognition and Tracking from Multi-Temporal Synthetic Aperture Radar and Radiometer Observations Rascle, Nicolas Gilles; Grouazel, Antoine; Mouche, Alexis; Nouguier, Frederic; Villarreal Olavarrieta, Carlos Eduardo; Mora Escalante, Rodney Eduardo; Diaz Mendez, Guillermo; Chapron, Bertrand; Ocampo-Torres, Francisco J. Satellite Measurement Of Waves And Currents: SAR Vs Optical Sensors Holt, Benjamin; Lockhart, Brittany; Porter, Mitchell; Comer, Douglas Fronts, Eddies, and Other Features in the Western Pacific and Micronesia from SAR and Other Multi-Sensor Data Husson, Romain; Ollivier, Annabelle; Chehade, Bassam; Peureux, Charles; Quet, Victor; Goimard, Gaël; Soulat, François; Tourain, Cédric; Lachiver, Jean-Michel Comparing and Combining S1 and SWIM Spectral Wave Measurements Methodology and Techniques Tings, Björn; Pleskachevsky, Andrey; Wiehle, Stefan; Jacobsen, Sven Ship Wake Detectability in TerraSAR X, CosmoSkymed, Sentinel 1 and RADARSAT 2 Imagery – Summary and Applications for Wake Detection Arthurs, David Open and Reproducible Science: The Role of Computing Platforms for Research and Applications that use SAR Gade, Martin; Peters, Sebastian; Schäfers, Simon On the Use of SAR Data to Monitor Coastal Erosion and Morphodynamics in Intertidal Areas Hajduch, Guillaume; Pinheiro, Muriel; Valentino, Antonio; Vincent, Pauline; Recchia, Andrea; Cotrufo, Alessandro; Franceschi, Niccolo; Piantanida, Ricardo; Benchaabane, Amine; Peureux, Charles; Husson, Romain; Schmidt, Kersten; Mouche, Alexis; Grouazel, Antoine; Nouguier, Frédéric; Johnsen, Harald; Hindberg, Heidi; Guiton, Gilles; Collard, Fabrice Sentinel-1 product performance Shamshiri, Roghayeh; Eide, Egil; Rangriz Rostami, Fazel; Vilhelm Høyland, Knut Sentinel-1 Extra Wide Thermal Noise Removal Using a Deep Learning Model Yitayew, Temesgen Gebrie; Grydeland, Tom; Larsen, Yngvar; Engen, Geir Processing of High Squint Bistatic SAR Data: The Case of Harmony Hoffman, Lauren Alexandra; Mazloff, Matt R; Gille, Sarah T; Giglio, Donata; Bitz, Cecilia M; Heimbach, Patrick Machine Learning for Evaluating the Drivers of Variability in Arctic Sea-ice Motion. Grydeland, Tom; Yitayew, Temesgen Gabrie; Larsen, Yngvar;DuBois, Pierre; Armstrong, Thomas; Gombert, Baptiste; Soulat, Francois; Monnier, Goulven; Hellouvry, Yann-Herve; Camus, Benjamin; Lopez-Dekker, Paco; Lajas, Dulce; Rommen, Bjorn; deWitte, Erik Signal Processing for Harmony: Illustrations with Simulated Data over Ocean Colin, Aurélien; Tandeo, Pierre; Husson, Romain; Fablet, Ronan; Peureux, Charles MediSAR: An Exhaustive Augmented Dataset Of Segmented Sentinel-1 SAR Ocean Observations Of The Mediterranean Sea and the Black Sea regions Stopa, Justin E.; Foster, Ralph; Vandemark, Doug; Wang, Chen; Chapman, Jonathan; Glaser, Yannik; Sadowski, Peter; Mouche, Alexis; Chapron, Bertrand Using Ocean Surface Imagery to Estimate Atmospheric Boundary Layer Stratification Foster, Ralph; Mouche, Alexis; Chapron, Betrand Using SAR Imagery to Diagnose Tropical Cyclone Boundary Layer Mean State Larsen, Yngvar; Engen, Geir; Grydeland, Tom; Yitayew, Temesgen G. An Alternative Approach for Estimation of Doppler Centroid Anomaly Based on Level-0 SAR Data Alpers, Werner R.; Bignami, Franceco Sar Observation Of Internal Waves Generated By Sub-mesoscale Features In The Strait of Sicily Applications: Maritime Security Frost, Anja; Kortum, Karl; Wiehle, Stefan; Tings, Björn. Ship Navigation Assistance For Polar Waters By Providing Information On Sea Ice Drift And Deformation Zones Using TerraSAR-X Data Yang, Yi-Jie; Singha, Suman); Goldman, Ron Integration of a Deep Learning Based Oil Spill Detection System into an Early Warning System for the Southeastern Mediterranean Sea Blondeau-Patissier, David; Schroeder, Thomas; Suresh, Gopika; Li, Zhibin; Diakogiannis, Foivos; Steven, Andrew Detection Of Marine Oil-like Features In Sentinel-1 SAR Images By Supplementary Use of Deep Learning and Empirical Methods McGourty, Sara; Kaczor, Scott; Capsey, Austin LR extrapolates much more successfully, if the modelled conditions are outside the training data conditions while the ML models can give an error with outliers significantly exceeding three times the RMSE. Further, in addition to ML training (can take months), the developed ML model is many orders of magnitude larger (takes Gigabytes) than the list of coefficients for the LR model (takes Kilobytes). LR outperforms ML in terms of parsing speed of the model, which is important for NRT services. This point is important, as a migration of the sea state processing for direct installation on a satellite for on-board-processing has been developed [10]. In this case, no huge amount of SAR raw data will be transferred from satellite to earth, before the processing can be done, but only already derived sea state parameters. This technology will significantly simplify the data transfer and reduce the delay. Based on all these reasons, the proposed SAR-SeaStaR algorithm combines both: LR (based on CWAVE approach [6] extended by series of additional features [9]) and ML model (using support vector machine (SVM) technique) for sea state processing. The solution of LR model (Hs) is used as first guess value for ML (additional feature) and also is applied for control of results. 2. Algorithm basic parameters In a classic way, the estimation of sea state parameters is based on a normalized radar crosssection (NRCS) analysis of subscenes. One of the basic variables represents the SAR image spectrum obtained using fast Fourier transformation FFT applied to the ground range detected, radiometrically calibrated, filtered, denoised land-masked and normalised subscenes with a size of 1,024×1,024 pixels in wave number domain as introduced in [5]. SAR features estimated from a subscene are of five different types: − NRCS and NRCS statistics (variance, skewness, kurtosis, etc.). − Geophysical parameters (wind speed using CMOD-5 algorithms for C-band [11] and XMOD-2 for X-Band [11]). − Grey Level Cooccurrence Matrix (GLCM) parameters (entropy, correlation, homogeneity, contrast, dissimilarity, energy, etc.). − Spectral parameters, based on image spectrum integration of different wavelength domains (030 m, 30-100 m, 100-400 m, etc.) and spectral width parameters (Longuet-Higgins, Goda). − Spectral parameters using products of normalized image spectrum with orthonormal functions (CWAVE approach) and cutoff wavelength estimated using autocorrelation function (ACF). 3. Sea state processor (SSP) SSP was designed in a modular architecture for S1 IW, EW, WV and TS-X SM/SL modes. The DLR Ground Station “Neustrelitz” applies the SSP as part of a near real-time demonstrator service that involves a fully automated daily provision of surface wind and sea state parameters estimated from S1 IW images of the North and Baltic Sea. Due to implemented parallelization, a fine raster can be processed. For example, S1 IW image with coverage of 200 km × 250 km can be processed using a raster with 1 km sized grid cells (~50,000 subscenes) during minutes. Each of maritime information products, i.e. sea state retrieval, wind speed retrieval, ship detection and AIS defines each one independent data layer. The data layers are combined for processor-internal information exchange and presentation to the operator [9]. Figure 1: Example of eight sea state fields processed from S-1 IW scene with ~1600 km by ~200 km coverage acquired during a storm in North Atlantic with Hs reaching ~14 m Using the SSP, the complete archive of S1 WV data from December 2014 until February 2021 with around 60 overflights/day, each including around 120 imagettes, was processed. The validation using the WFWAM/CMEMS [2] model resulted in an RMSE of 0.245/0.273 m for wv1/wv2 imagettes, respectively. Comparisons to 61 NDBC buoys [13], collocated at distances shorter than 50 km to worldwide S1 WV imagettes, result into an RMSE of 0.41 m. The data is available to the public within the scope of ESA’s climate change initiative CCI [14]. References [1] SARWAVE. www.sarwave.org [2] CMEMS. Copernicus Marine Environment Monitoring Service. https://marine.copernicus.eu/ [3] Schwarz, E., Krause, D., Berg, M., Daedelow, and H. Maas, 2015. Near Real Time Applications for Maritime Situational Awareness. Proc. 36th Int. Symp. On Remote Sensing of Environment, Berlin, Germany, Remote Sensing and Spatial Information Sciences, vol. XL-7/W3, 6p. [4] Pleskachevsky, A., W. Rosenthal, and S. Lehner, 2016: Meteo-Marine Parameters for Highly Variable Environment in Coastal Regions from Satellite Radar Images.” JPRS, Vol. 119, pp. 464-484., 2016. [5] Pleskachevsky, A., Jacobsen, S., Tings, B., and E. Schwarz, 2019. Estimation of sea state from Sentinel-1 Synthetic aperture radar imagery for maritime situation awareness. IJRS, Vol. 40-11, pp. 4104-4142. [6] Schulz-Stellenfleth, J., König, Th., and S. Lehner, 2007. An empirical approach for the retrieval of integral ocean wave parameters from SAR data. JRL, Vol. 112, pp. 1-14. [7] Stopa, J., and A. Mouche, 2017. Significant wave heights from Sentinel-1 SAR: Validation and Applications. JGR, Vol., 122, pp. 1827-1848.. [8] Quach, B., Glaser, Y., Stopa, J., Mouche, A., Sadowski, P., 2020. Deep Learning for Predicting Significant Wave Height From Synthetic Aperture Radar, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, 10.1109/TGRS.2020.3003839 [9] Pleskachevsky, A., Tings, B., and J, Jacobsen, 2022: Multiparametric Sea State Fields from Synthetic Aperture Radar for Maritime Situational Awareness, RSE, vol. 280, 22 pp. [10] Wiehle, S., Breit, H., Günzel., D., Mandapati, S., and U. Balss, 2022. Synthetic Aperture Radar Image Formation and Processing on an MPSoC, IEEE Transactions on Geoscience and Remote Sensing, DOI: 10.1109/TGRS.2022.3167724 [11] Hersbach, H., 2008. CMOD5.N: A C-band geophysical model function for equivalent neutral wind. [12] Li, X. -M. and S. Lehner, 2014. Algorithm for Sea Surface Wind Retrieval From TerraSAR-X and TanDEMX Data, IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 5, pp. 2928-2939. [13] https://www.ndbc.noaa.gov/ [14] ESA, CCI, Sea State Project. http://cci.esa.int/seastate SAR-altimetry cross-spectra M. Kleinherenbrink, F. Ehlers, F. Gibert, S. Hernandez, F. Nouguier, B. Chapron, P. Lopez-Dekker January 31, 2023 1 Introduction The normalized radar cross section (NRCS) from Synthetic Aperture Radar (SAR) depends on the short-scale roughness of the ocean surface[10]. Modulations of the NRCS result from the presence of long gravity waves, such as swell. Since the 1970s these intensity modulations have been studied to infer wave parameters. The mapping of wave spectra into SAR intensity spectra[5, 8] led to the first retrieval algorithms[2], which are now operational for SAR satellite missions such as Sentinel-1. Distortions of intensity by long waves have also been observed in high-resolution delay/Doppler altimetry, which affects the accuracy to which geophysical parameters can be retrieved[9]. The limited along-track resolution prevents the interpretation of swell-related waveform distortions. With Focused SAR-altimetry processing, which is applicable to delay/Doppler altimeters, it is possible to compute SAR spectra from distortions in the waveform tails in a comparable way as the side-looking SAR systems. Instead of two ambiguities in side-looking SAR[4], the SAR-altimetry spectra contain four ambiguities because both sides of the ground track are illuminated[1]. Under moderate sea states it should be possible to retrieve swell-wave parameters from nadir altimeters. However, an inversion or retrieval of swell-wave parameters has not been performed yet as the SAR-altimetry spectra are still poorly understood. This paper will describe SAR-altimetry spectra in more detail using a closed-form solution and simple model simulations, which are compared to Sentinel-6 data. It extends the description of Altiparmaki et al.[1] and introduces the concept of the cross-spectral stack, a set of cross-spectra using multiple sublooks from the overpass. We will show that the cross-spectra vary during the overpass, which can help to constrain the estimation of swell-wave parameters. 2 Closed-form model The closed-form model [5, 8], suffers from limitations and is not accurate for the steep incident angles of nadir altimetry (more details in Kleinherenbrink et al. (2023), in preparation). However, the closed-form model can provide a local approximation for the observed spectrum and helps to understand the SARaltimetry spectral properties. We define the two-dimensional SAR spectrum for the right side of the ground-track as [4] P(kx, ky) = 1 (2π)2e−k2 xρxx (0,0)−k2 yρyy (0,0)−kxky(ρxy (0,0)+ρyx (0,0)) Z Z ek2 xρxx +k2 yρyy +kxky(ρxy +ρyx )Le−i(kxx+kyy)dxdy, (1) with L= 1 + ρII +... ρab cross-correlation functions computed from ocean-wave spectrum S(kx, ky) and transfer functions for ground-range shifts, Doppler shifts and tilt modulation as [6, 4] Tx=−1 tan(θ) Ty=−iR Vωk TI=−ikx 1 σ0 δσ0 δθ , (2) where θis the local incident angle, R/V the reciprocal of the angular satellite velocity, ωkthe angular velocity of the ocean waves and σ0the NRCS. This formulation explicitly takes into account the non-linear range shifts (range bunching). The exponential terms outside of the integral provide us a clue on the resolution in the alongand cross-track directions, approximated as λal ∝πqρyy(0,0) = πH Vpσ2 v λct ∝πpρxx(0,0) = πsσ2 e tan2(θ)=πSWH 4 tan(θ), (3) with σ2 vthe velocity variance [7], and the Significant Wave Height (SWH) as a function of the elevation variance σ2 e. This yields an ellipsoidal roll-off in the spectral domain. In case the range direction and the azimuth direction are not perpendicular, the orientation and size of the ellipse changes. 3 Numerical model A surface is modelled on a regular grid using as input an Elfouhaily wind-wave spectrum[3] and a Gaussian swell-wave spectrum. Using Discrete Fourier Transforms we compute the elevation h, upward velocity vrand two-dimensional slope. The local in-plane and out-of-plane incident angles are used to compute the tilt modulation using the well-known equation for specular reflections[10]. For the zero-Doppler case, the surface velocity is used to reproject scatterers in the Doppler along-track direction ∆y=R Vvr(4) Figure 1: Absolute values of numerically modeled satellite nadir altimetry cross-spectra based on ten realizations at five alongtrack distances. The inputs consist of swell systems with a peak wavelength of 277 m propagating at 45 (top) and 135 (bottom) degrees from cross track (right), and a wind-sea system based on a wind speed of 10 m/s, a fetch of 200 km at a mean direction of 45 degrees from cross track. Figure 2: Imaginary values of numerically modeled satellite nadir altimetry cross-spectra based on ten realizations at five alongtrack distances. The input consists of a swell system with a peak wavelength of 277 m propagating at 45 (top) and 225 (bottom) degrees from cross track (right), and a wind-sea system based on a wind speed of 10 m/s, a fetch of 200 km at a mean direction of 45 degrees from cross track. and the elevation to move scatterers is the ground-range (crosstrack) direction by ∆x=−1 tan(θ).(5) The latter two effects are responsible for the resolution losses in the along-track and the cross-track direction, respectively, and also for the accompanied velocity and range bunching. 4 Cross-spectral stack The 3dB-beamwidth of radar altimeters is typically in the order of 1.3 degrees. At a platform velocity of approximately 7 km s−1scatterers are a few seconds in view. Only using part of the overpass for SAR processing is sufficient to get the necessary along-track resolution for spectral analysis. By dividing the overpass in multiple sublooks, we get observations and spectra of the surface from various view angles leading to varying spectral densities. Cross-spectra are computed from two consecutive sublooks [4], which provides additional information on the phase change of the intensity modulations. Without considering the phase, the history of the spectral density emerging from a swell signal varies by four effects. First, the tilt modulation varies as the relative satellite position with respect to the slopes changes. Second, the ground-range direction is at non-zero Doppler not aligned anymore with the crosstrack direction. Surface elevations still cause reprojections of scatterers in the cross-track direction, but their effects are ’filtered’ and depend on the width of the Doppler strip, which leads to an adjustment of Eqs. 2 and 5. Third, vertical velocity causes Doppler shifts that move scatterers in the along-track direction. However, as they are constrained by their range, reprojection occurs effectively over the range isoline. This introduces a cross-track shift and adds a term to the ground-range transfer function Tx(Eq. 2) and to the shift in Eq. 5. Fourth, the range and Doppler shifts are not perpendicular anymore, which causes the cross-term outside of the integral (Eq. 1) to becomes non-zero. This will rotate the cut-off and lowers the power spectral density in two quadrants. The combined effect results in different power spectral density variations in the spectral quadrants (Fig. 1). It appears that the power spectral density in a non-zero Doppler crossspectrum depends on the wind-wave and swell directions. The imaginary terms of the cross-spectra also vary during the overpass (Fig. 2). The phases or imaginary parts of the crossspectra cannot directly be coupled to the wave direction, as the location of maximum intensity also varies due to the changing phases of the three imaging mechanisms by the geometrical variations induced by the satellite velocity. The relative phase change of the ’static’ intensity modulation (comparable to SAR, see Nouguier et al. (2023), in preparation) interferes with the phase change caused by wave motion. The cross-spectral stack might therefore be used to remove ambiguities, but it is nontrivial and the precise mechanism is not fully understood yet. 5 Conclusions This study provides the most extensive description of the SARaltimetry spectra up to date. It also introduces the crossspectral analysis for SAR altimeters. The long overpass time of the altimeters allows to estimate a cross-spectral stack. The cross-spectral stack adds information for retrieval algorithms, but is not yet fully understood. It also shows potential to remove swell-direction ambiguities. The next step is the inversion of swell-wave parameters from the SAR-altimetry cross-spectra. If swell-wave parameters are retrieved from altimeters this will greatly increase the number of ocean-wave observations. In combination with the backscatter (roughness), the SWH and the velocity variance, the SAR altimeter will have a more synoptic view of the ocean surface. Finally, the additional information resulting from such an inversion would help to constrain the sea-state bias. References [1] O. Altiparmaki, M. Kleinherenbrink, M. Naeije, C. Slobbe, and P. Visser. SAR Altimetry Data as a New Source for Swell Monitoring. Geophysical Research Letters, 49(7), Apr. 2022. [2] B. Chapron, H. Johnsen, and R. Garello. Wave and wind retrieval from sar images of the ocean. Annales Des T´el´ecommunications, 56(11-12):682–699, Nov. 2001. [3] T. Elfouhaily, B. Chapron, K. Katsaros, and D. Vandemark. A unified directional spectrum for long and short wind-driven waves. Journal of Geophysical Research: Oceans, 102(C7):15781–15796, July 1997. [4] G. Engen and H. Johnsen. SAR-ocean wave inversion using image cross spectra. IEEE Transactions on Geoscience and Remote Sensing, 33(4):1047–1056, July 1995. [5] K. Hasselmann and S. Hasselmann. On the nonlinear mapping of an ocean wave spectrum into a synthetic aperture radar image spectrum and its inversion. Journal of Geophysical Research, 96(C6):10713, 1991. [6] F. C. Jackson. THE physical basis for estimating waveenergy spectra with the radar ocean-wave spectrometer. Johns Hopkins APL Technical Digest, 8(1), 1987. [7] V. Kerbaol, B. Chapron, and P. W. Vachon. Analysis of ERS-1/2 synthetic aperture radar wave mode imagettes. Journal of Geophysical Research: Oceans, 103(C4):7833– 7846, Apr. 1998. [8] H. E. Krogstad. A simple derivation of Hasselmann’s nonlinear ocean-synthetic aperture radar transform. Journal of Geophysical Research, 97(C2):2421, 1992. [9] T. Moreau, E. Cadier, F. Boy, J. Aublanc, P. Rieu, M. Raynal, S. Labroue, P. Thibaut, G. Dibarboure, N. Picot, L. Phalippou, F. Demeestere, F. Borde, and C. Mavrocordatos. High-performance altimeter Doppler processing for measuring sea level height under varying sea state conditions. Advances in Space Research, 67(6):1870–1886, Mar. 2021. [10] R. Valenzuela. Theories for the Interaction of Electromagnetic and Oceanic Waves - A Review. Boundary-Layer Meteorology, 13:61–85, 1978. Wave Density Spectra in the Baltic Sea from Sentinel-1 SAR Imagery Using the LSTM Model Sander Rikka 1, Sven N˜omm2, Victor Alari1, Jan-Victor Bj¨orkqvist3,4, Martin Simon2 1Department of Marine Systems, School of Science, Tallinn University of Technology TalTech, Akadeemia tee 15a, 12618, Tallinn, Estonia, [email protected] 2Department of Software Science, School of Information Technology, Tallinn University of Technology, Akadeemia tee 15 a, 12618, Tallinn, Estonia, 3Norwegian Meteorological Institute, 5007, Bergen, Norway. 4Finnish Meteorological Institute, 00560, Helsinki, Finland, Abstract This paper presents preliminary results on the use of a long-short-term memory (LSTM) deep recurrent neural network to estimate 1D wave density spectra from Sentinel-1 (S1) Interferometric Wide (IW) swath images. In total, 165 locations in the Baltic Sea are used to extract image spectra from custom ground-range detected data. The training data set consists of approximately 80000 individual data points divided into training, validation, and testing data sets by 70-15-15, respectively. WAM wave model spectra from the MET Norway repository are used as ground truth. The LSTM model shows correlations greater than 0.80 between wave periods of 2 and 9 seconds for the wave density spectra. 1. Introduction Estimation of various surface wave parameters (e.g. significant wave height HS) from SAR imagery has been the focus of many studies [1], [2] and [3]. However, the accurate retrieval of 2D wave density spectra has been the ultimate goal of the SAR data. Various inverse methods have been developed to extract 2D wave information for the open ocean [4], [5] where high resolution wave mode data are acquired and swell waves dominate the sea state. The lack of long-swell waves in closed wind-wave dominant regional seas, e.g. the Baltic Sea [6], [7], and lower-resolution ScanSAR (TOPS mode for S1) acquisitions have been the limiting factor for using the inverse methods for such environments. Figure 1: Map of the Baltic Sea with locations of SAR data collection. Colors represent the amounts of SAR match-ups with model spectra. Figure 2: NORA3 significant wave height distribution over mean propagation directions (going to) for the matching dataset. Recently, recurrent deep learning neural network methods (e.g., long-short-term memory - LSTM [8]) have been used to estimate 1D wave density spectra from S1 IW data in the Baltic Sea [9]. The aim of this study is to extend 1 the data set by using NORA3 wave model spectra and train LSTM networks to estimate wave density spectra between 2 and 10 s with a much larger and more variable data set compared to the previous study. 2. Data and Methods 2.1 NORA3 data We used the NORA3 wave hindcast, which is based on the WAM wave model [10], [11]. The hindcast covers the pan-Arctic domain, including also the Baltic Sea with a 3 km resolution. However, the spectra were only outputted with a 30 km resolution. The model spectra are discretized using 30 frequencies logarithmically spaced in the range of 0.0345-0.5476 Hz and 24 equally spaced directions covering a full circle. 2.2 SAR processing S1 IW Single Look Complex (SLC) sub-images were calibrated and speckle filtered with a Frost (5x5) filter. During the multi-look operator, pixels were not averaged to represent squares, and the images were left in radar projection. The image spectra (ISP ) are calculated from both polarisations (V V and V H) with the fast Fourier transform (FFT). Subsequently, the values of the ISP components ISPxand ISPywere interpolated to fixed wavelength values between 215 and 30 m (39 values with variable intervals). Image spectra and other metadata (e.g. satellite heading (P ASS), incidence angle (IA), image texture, etc.) were saved for later processing. SAR data was collected around the location of the model spectra (Figure 1) from the beginning of 2015 until the end of 2021. On average, around 550 independent sub-images were processed at each location. The resulting distribution of HSin the mean propagation direction is shown in Figure 2. In total, around 80000 collocation pairs were formed for the study. 2.3 LSTM configuration and experiments Spectra estimation was performed by a deep LSTM type neural network. The network architecture is shown in Figure 3. The best configuration was found when the input data for the LSTM model had a shape of (40,4), where the four variables are IA and ISPV V x,P ASS and ISPV V y,IA and ISPV Hx,P ASS and ISPV Hy. Before training, the values IA and P ASS are normalized between 0 and 1; the logarithm base eis applied to ISP and wave energy. 3. Results and discussion Bin-by-bin comparison in Figure 4 between test data and LSTM model predictions show correlations greater than 0.80 for all predicted spectra between 2 and 9 s, which corresponds to most of the sea states dominated by wind waves. Although the correlation coefficient for the energy density around 10 s in [9] is greater, the prediction accuracies for higher frequencies are similar or even better for the current study. Moreover, the custom SAR processing (without squaring pixels during multi-look) allows for avoiding complex image upsampling which was required previously to estimate energies in higher frequencies. Figure 3: LSTM model structure. Figure 4: LSTM model prediction average correlation coefficients on test data for wave energy per frequency/period band and standard deviation of errors. 2 Acknowledgements We thank MET Norway for its open data policy and data availability. The corresponding author expresses sincere gratitude to the co-authors. The work of S. N˜omm and M. Simon is in the project ”ICT programme” which was partially supported by the European Union through the European Social Fund. References [1] S. Lehner, A. Pleskachevsky, D. Velotto, and S. Jacobsen, “Meteo-marine parameters and their variability: Observed by high-resolution satellite radar images,” Oceanography, vol. 26, no. 2, pp. 80–91, 2013. [2] F. Ardhuin, J. E. Stopa, B. Chapron, F. Collard, R. Husson, R. E. Jensen, J. Johannessen, A. Mouche, M. Passaro, G. D. Quartly et al., “Observing sea states,” Frontiers in Marine Science, p. 124, 2019. [3] A. Pleskachevsky, B. Tings, S. Wiehle, J. Imber, and S. Jacobsen, “Multiparametric sea state fields from synthetic aperture radar for maritime situational awareness,” Remote Sensing of Environment, vol. 280, p. 113200, 2022. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0034425722003108 [4] K. Hasselmann and S. Hasselmann, “On the nonlinear mapping of an ocean wave spectrum into a synthetic aperture radar image spectrum and its inversion,” Journal of Geophysical Research: Oceans, vol. 96, no. C6, pp. 10 713–10 729, 1991. [5] J. Schulz-Stellenfleth, S. Lehner, and D. Hoja, “A parametric scheme for the retrieval of two-dimensional ocean wave spectra from synthetic aperture radar look cross spectra,” Journal of Geophysical Research: Oceans, vol. 110, no. C5, 2005. [6] J.-V. Bj¨orkqvist, S. P¨art, V. Alari, S. Rikka, E. Lindgren, and L. Tuomi, “Swell hindcast statistics for the baltic sea,” Ocean Science, vol. 17, no. 6, pp. 1815–1829, 2021. [Online]. Available: https://os.copernicus.org/articles/17/1815/2021/ [7] L. Tuomi, K. K. Kahma, and H. Pettersson, “Wave hindcast statistics in the seasonally ice-covered Baltic Sea,” Boreal Environ. Res., vol. 16, no. 6, pp. 451–472, 2011. [8] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 11 1997. [Online]. Available: https://doi.org/10.1162/neco.1997.9.8.1735 [9] M. Simon, S. Rikka, S. N˜omm, and V. Alari, “Application of the lstm models for baltic sea wave spectra estimation,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 83–88, 2023. [10] Ø. Breivik, A. Carrasco, H. Haakenstad, O. J. Aarnes, A. Behrens, J.-R. Bidlot, J.-V. Bj¨orkqvist, P. Bohlinger, B. R. Furevik, J. Staneva, and M. Reistad, “The impact of a reduced high-wind charnock parameter on wave growth with application to the north sea, the norwegian sea, and the arctic ocean,” Journal of Geophysical Research: Oceans, vol. 127, no. 3, mar 2022. [Online]. Available: https://doi.org/10.1029\%2F2021jc018196 [11] “Norwegian meteorological institute, spectra database,” https://thredds.met.no/thredds/catalog/windsurfer/ mywavewam3km spectra/catalog.html, accessed: 2022. 3 Temporal Consistency of Sentinel-1 SAR Wave Mode Level-2 Ocean Swell Spectra Salman Khan Commonwealth Scientific and Industrial Research Organisation, Environment, Aspendale, Victoria, Australia. Sentinel-1 C-SAR platforms (A/B) acquire small wave mode images covering most of the global ocean up to ~100 km close to the coast [1]. Ocean surface wave information is routinely extracted from these images such as in the ESA Sentinel-1 level2 ocean swell (OSW) spectra product [2]. The swell spectra derived from SAR (including Sentinel-1) are limited by the azimuth cut off wavelength and generally cannot capture high frequency ocean waves. Despite this limitation, they are useful in monitoring and analysing longer scale swell systems, and prior to CFOSat’s (China France Oceanography satellite) launch, they were the only satellite-derived ocean swell spectra measurements that were routinely produced. The algorithm that sits behind the production of the OSW product has matured over time through several updates [3,4]. It would be ideal to reprocess the whole Sentinel-1 wave mode archive at each OSW processor update to have a product that is the most accurate and consistent in time, but that is not currently the case and may be too costly a prospect. Nevertheless, it is important to examine the temporal consistency of the OSW product before potentially using these data for various applications. In this brief work, a temporal consistency analysis of the OSW product has been carried out in the Australian region. In this region, there are not many wave buoy measurements as one moves further away from the coast. Therefore, one viable option for a mission duration temporal consistency analysis is through comparisons with regional wave model runs. We have taken this approach and have carried out the comparison of the OSW product against an Australian-produced global wave model (WaveWatch III) hindcast [5]. The matching criteria between OSW observations and the wave hindcast are within 100 km and +/- 30 mins. The wave spectra from the hindcast are also truncated using the ellipsoidal azimuth cut off from the matching OSW spectra as done in [6]. The comparison is performed using bulk statistics of significant wave height, zero-crossing period, and mean direction (Hs, tm02, and dm) to assess the consistency of wave parameters over time. In each comparison, the time series of the mean monthly bias and standard deviation (Sentinel-1 - WW3) are analysed after grouping the data by platform, polarisation, incidence angle, and Sentinel-1 IPF (Instrument Processing Facility) versions [4]. This approach gives clues into the dependency of the wave parameters/swell spectra on the above characteristics and the processor version. The results for the three cases are shown in figs. 1-3. The results clearly show that the biases and standard deviations are reduced with improvement to the Sentinel-1 IPFs generally for all combinations of satellite characteristics with only some exceptions. Especially, the results since IPF 3.30+ are quite encouraging. Resources permitting, there is value in reprocessing historical Sentinel-1 wave mode data with the latest IPF version. Until that is possible, the users of this dataset need to consider its limitations due to evolving IPF versions before their use. A subsequent analysis on potential calibration consistency of the Sentinel-1 affecting wave mode data was also carried out. The calibration consistency was analysed on the normalised radar cross section (NRCS) time series. The NRCS is a function of incidence angle, operating frequency, and transmit-receive polarisation. Daily NRCS mean and standard deviation on Sentinel-1 wave mode time series were computed to check for inconsistencies after grouping the data by platform, incidence angle, polarisation and IPF version (fig. 4). The NRCS stats of Sentinel-1 A and B wave mode level-2 data for overlapping periods, matching polarisation and incidence angles closely follow each other and no alarming inconsistency is observed. Slightly higher fluctuation of mean NRCS value is possibly observed in the beginning of the time series (IPF 2.50, fig. 4). Figure 1: (left y-axis) ∆𝐻𝑠 (Sentinel-1 – WW3) monthly mean and standard deviation time series of Sentinel-1A/B wave mode data as a function of platform, polarisation, and incidence angle, and (right y-axis) WW3 monthly mean 𝐻𝑠 computed against Sentinel-1 A A and B matchups. Sentinel-1 IPF version evolution through time is highlighted. Figure 2: Figure 7: (left axis) ∆𝑡𝑚02 (Sentinel-1 – WW3) monthly mean and standard deviation time series of Sentinel1A/B wave mode data as a function of platform, polarisation, and incidence angle, and (right axis) WW3 monthly mean 𝑡𝑚02 computed against Sentinel-1 A and B matchups. Sentinel-1 IPF version evolution through time is highlighted. Figure 1: Density map for co-located data between the WV1 SAR acquisitions and WW3 over the period from July 2021 to August 2022. Figure 1 illustrates the density maps of collocated data for WV1 (similar for WV2). The full dataset includes more than 348 000 pairs. For this study, several variables were extracted from ESACopernicus WV level 2 (L2 OCN) products identified as potential features that are expected to explain the wind-sea significant wave height variability. These variables include both geophysical information such as wind speed and direction and other acquisition information like incidence angle and azimuth cut-off. The list of these variables is shown in Table 1. A whole feature engineering pipeline has been put in place to prepare data for learning. As for the target, it is defined as the wind-sea Hs estimated from the WW3 spectra. To get such variables, the first step is to extract the wind sea spectra from the WW3 spectra collocated with the SAR, then we estimate the associated Hs. The distribution of this computed variable, presented by a range of Hs for both collocated WV1 and WV2 dataset, is shown in Figure 2. The obtained distribution of wind-sea Hs is not uniform and contains extreme values that are poorly represented. Initially, this database is used to define a baseline of learning performances. Then, it is than submitted to a specific balancing task to obtain a more homogeneous distribution of wind-sea Hs and to perform model accuracy. As for each machine learning model, the database was divided into three parts: a training data set (~122200 pairs), a validation data set (~26000 pairs) and a test data set (~26000 pairs). b. Method The first tests are performed using a simple deep learning architecture defined by a Deep Networks Regressor (DNNRegressor). The SAR features are fed into 5 dense layers of 128,64,32,16,8 units respectively with Rectified Linear Unit (ReLU) as an activation function. Between each two hidden layers a Batch Normalization was applied to perform training. Finally, the wind-sea Hs prediction is performed by an output layer with linear activation function. This model is trained to minimize the mean squared error (MSE) using Adam optimizer with a batch size of 150 samples. A specific learning rate decay was applied to improve the optimization. Training was stopped when the validation loss did not improve after 10 epochs to avoid overfitting. The architecture, learning rate, early stopping and batch size was optimized by tuning and only for WV1. The same hyperparameters were used to train WV2 model to investigate the incidence angle sensitivity. SAR Features Signal to Noise Ratio (SNR) Normalized Radar Cross Section (NRCS) Azimuth Cut-off SAR wind speed SAR wind direction in sensor convention SAR wave spectrum Kurtosis SAR wave spectrum Skewness SAR wave normalized variance Incidence angle of acquisition Table 1: SAR features used as inputs of the DNN-Regressor. Figure 2 : Wind-sea Hs count by 0.5m bins for WV1 and WV2. Metric WV1 WV2 Bias -0.02 m -0.04 m Standard deviation 0.49 m 0.45 m Correlation 0.92 0.93 R2 score 0.83 0.86 RMSE 0.49 m 0.45 m Table 2: Statistics of model’s performances for WV1 and WV2. 3. Results and perspectives a. Results The first results obtained are very promising and the two models obtained for WV1 and WV2 generalize well on the test data (Figure 3). The performances indicators are displayed in Table 2. First, we notice that the models suffer a little from high wind-sea Hs values by underestimated them. This is starting to appear from Hs > 4m for WV1 and Hs > 2m for WV2. This finding is expected because this model has not been tuned as the case of WV1. Therefore, this confirms implicitly that there is a sensitivity to incidence and that the two modes should be treated separately since the acquisition noise differs according to incidence. The extreme wind-sea Hs under-estimation is also explainable since they are poorly represented in the training distribution as introduced earlier. Work is in progress to balance the training data set for all ranges of windsea Hs. b. Perspectives We have created a dataset of over 348000 collocations between SAR features, derived wind-sea Hs from WW3 and used it to train a DNN-Regressor model that predicts wind-sea Hs from SAR features. The obtained models underestimate extreme Hs which is tackled by a synthetic data generation to balance the training dataset. Figure 4 illustrates the generative potential of extreme Hs for the case of WV1 under validation: red bars indicate synthetic data; blue bars indicate initial data; green bars indicate assembled data and gray bars for rate of generation. To improve the performance of the two models, new variables will be introduced such as IMACS [5]. Then, with this new enriched and balanced database, the architectures will be redesigned to consider the sparse and dense nature of each learning SAR feature. The relative contribution of these chosen input features will be investigated to quantify their relative importance. The estimation of a SAR-derived windsea component, together with the already existing swell and “total significant wave height” paves the way for complete SAR-derived directional wave spectra. 4. References [1] Engen, G., and H. Johnsen. “SAR-Ocean Wave Inversion Using Image Cross Spectra.” IEEE Transactions on Geoscience and Remote Sensing 33, no. 4 (1995): 1047–56. [2] Li, X. M., S. Lehner, and M. X. He. “Ocean Wave Measurements Based on Satellite Synthetic Aperture Radar (SAR) and Numerical Wave Model (WAM) Data – Extreme Sea State and Cross Sea Analysis.” International Journal of Remote Sensing 29, no. 21 (2008): 6403–16. . [3] Pleskachevsky, Andrey, Sven Jacobsen, Björn Tings, and Egbert Schwarz. “Estimation of Sea State from Sentinel-1 Synthetic Aperture Radar Imagery for Maritime Situation Awareness.” International Journal of Remote Sensing 40, no. 11 (June 3, 2019): 4104–42. https://doi.org/10.1080/01431161.2018.1558377. [4] Quach, Brandon, Yannik Glaser, Justin Edward Stopa, Alexis Aurélien Mouche, and Peter Sadowski. “Deep Learning for Predicting Significant Wave Height From Synthetic Aperture Radar.” IEEE Transactions on Geoscience and Remote Sensing, 2020. [5] Nilsen, Vegard, Geir Engen, and Harald Johnsen. “A Novel Approach to SAR Ocean Wind Retrieval.” IEEE Transactions on Geoscience and Remote Sensing PP (May 6, 2019): 1–10. https://doi.org/10.1109/TGRS.2019.2909838. [6] Li, H., Chapron, B., Mouche, A. A., & Stopa, J. E. (2019). A new ocean SAR cross-spectral parameter: Definition and directional property using the global Sentinel-1 measurements. Journal of Geophysical Research: Oceans, 124, 1566–1577. https://doi.org/10.1029/2018JC014638. [7] Johnsen, Harald, Romain Husson, Benchaabane Amine, Pauline Vincent, and Guillaume Hajduch. “Sentinel-1 Ocean Swell Wave Spectra (OSW) ATBD - Sentinel Online.”, S1-TN-NRT-52-7450, v1.5, 10 October 2022. shorturl.at/pqxX6 Figure 3: Scatterplot of wind-sea Hs estimated by the DNN-Regressor model against the WW3 target. Figure 4: Histogram of wind-sea Hs for initial and synthetic datasets. Ocean wave mapping by SAR: Numerical and theoretical perspectives Fr´ed´eric Nouguier1∗, Alexis Mouche1, Bertrand Chapron1, Marcel Kleinherenbrink3 1 Introduction Hasselmann and Hasselmann [2] and Krogstadt [3] paved the way of understanding the SAR imaging process of ocean waves by deriving the basic theoretical equations. Later on, Engen [1] reformulated them using characteristic functions and provided a stand alone methodology to remove wave propagation ambiguity using the evaluation of cross-spectra analysis between sublooks in azimuth. This technique elaborated during ESA ERS time was then applied on ENVISAT - ASAR acquisitions and is still currently applied for wave retrieval on Sentinel-1 acquisition over ocean without major evolution since the first derivation of the algorithm. Ocean wave spectrum retrieval from SAR observation remains a challenge due to its complex imaging mechanism and existing operational algorithms mainly rely on the exploitation of the so-called theoretical close-form. This close-form equation is a theoretical equation mapping the ocean wave spectrum into the back-scattered intensity SAR spectrum. Even if this equation exhibits most of the SAR imaging non-linear aspects (velocity bunching, azimuthal cut-off, ...), it still misses some key features that have not been correctly understood up to now. One striking example is that cross-spectra phase predicted by theoretical approach should be close to the waves dispersion relationship which is the dominant apparent displacement on the surface. Data analysis reveals that this wave dispersion relationship is strongly altered by non-linear imaging process of the SAR and can not be correctly predicted with the currently used close-form. Since the first attempts to derive wave spectrum from SAR observations, new techniques have been developed to model and understand the complex imaging process of waves by SAR. Computational capabilities have strongly evolved and opened new perspectives in simulating the complete chain of a SAR acquisitions on realistic synthetic ocean scene. Coupled with a revisit of the assumptions of the close-form derivations and foreseen neural-network methodologies, the authors developed a coherent framework to better understand and model ocean wave spectrum mapping by existing Synthetic Aperture Radar systems and support the development of new concept missions. To exemplify new perspectives offered by the aforementioned techniques, the authors selected some examples and the corresponding illustrations where they reveal themselves to provide interesting insights on the imaging process. In the coming years, it is planned to revisit the existing Sentinel-1 wave inversion algorithms and take benefit of these new techniques to better understand wave mapping by new coming SAR missions such as Harmony constellation based on innovative bistatic SAR configuration or Rose-L SAR operating at L band frequency. 2 The case of reversed swell propagation direction on Sentinel-1 Since the pioneer work of Engen et al. [1], it is commonly accepted that the sign of the imaginary part of the cross-spectrum computed between two looks extracted from Single Look Complex SAR data is directly linked to the ocean wave direction propagation. The wave displacement between the two looks (different acquisition time) is indeed assumed to be responsible for the non-vanishing phase of cross-spectrum between looks extracted from the azimuth Doppler Bandwidth. The resulting sign of the complex cross-spectrum is thus usually used to remove the ambiguity of the wave direction propagation in wave retrieval algorithms. Figure 1 shows a succession of six Sentinel-1 vignettes (WM1) acquisitions on July 6th 2020 between 15h08m14s and 15h10m40s in the Indian ocean where a long swell is captured by the sensor. In this specific case, the swell propagates towards the East direction. In this low latitudes part of Sentinel-1 orbit, the SAR range direction is rapidly changing relatively to the swell direction of propagation. While the swell direction is located slightly below the range direction on the first vignette (horizontal axis), it is located above the range axis on the last acquired vignette with a continuous variation during the intermediate acquisitions. Figures 2 show the imaginary part of the cross-spectrum for each acquisition. As observed, the sign of the swell peak energy is changing along the orbit (brown and purple colors are of opposite signs) while a constant positive sign is expected due to the established easterly swell system. At the middle acquisition (vignette #4 - 15h09m42s), the energy of the cross-spectrum is even split by the range axis in two parts with opposite signs. The theoretical close-form equation is unable to reproduce the change of sign in the cross-spectrum revealing some unexpected limitation of the theoretical approach. Several legitimate assumptions could be raised to explain the observed discrepancies such as the inaccuracy of the wave model compared to the true ground truth or the limiting assumption of the analytical approach such as the non-Gaussian nature of the ocean waves or a different look energy weighting leading to an inaccurate estimation of the separation time between looks. However, for the presented case, none of these assumptions explain the discrepancies and will be illustrated hereafter. In real data analysis, it is usually very hard to distinguish the reasons of this apparent discrepancies between the SAR measurements and the sea-state conditions. On the contrary, numerical simulations can be of great help in doing such parameters analysis thanks to the ability of disabling other possible effects. In this case, we relied on the “Remote Sensing Simulation from Space” (R3S) numerical model to explore the possible parameters that could enter into the aforementioned problematic case. This numerical model is a brute force simulation Figure 1: Succession of six Sentinel-1A vignettes (WM1) acquired in indian ocean. Vignette #1 was acquired at latitude -56.54 degree and vignette #6 at -48.4 deg. The same long swell (≈500m) is observed on all vignettes and its direction is slowing crossing the SAR range direction (horizontal axis). Figure 2: Imaginary part of the cross-spectra (between looks). Brown and purple colors are of opposite sign.. The sign of the energy corresponding to the observed swell around 500 m wavelength is changing from vignette #1 to vignette #6. Cross-spectrum #4 exhibits an energy blob splitted around the range axis with parts of opposite signs. Figure 3: R3S numerical simulation. Left panel is synthetic Sentinel-1 vignette corresponding to vignette #4 of figure 1 and rigth panel is the corresponding imaginary part of cross-spectrum. On both real and simulated cross-spectrum, we can observe the energy slpitting over the range axis with opposite signs on each side of it. intending to model as much physically as possible all the elements of the SAR observation chain of an ocean scene (ocean wave properties, scattering mechanisms of rough surface, satellite platform dynamics, antenna, SAR processing, ...). Figure 3, left panel, shows the simulated imagette which was generated using the co-localized Wave Watch III wave model and Sentinel-1 configuration corresponding to the acquisition date of July 6th 2020 15h0942 (vignette #4). The resulting imaginary cross-spectrum (right panel) exhibits the same sign inversion and splitting around the range axis on the simulated data supporting the fact that non-simulated effects (such as nonlinear wave statistics, ...) are not responsible of it. To push further the analysis, another simulation (not shown here) was processed but with a “frozen swell” meaning that the swell propagation was disabled during the SAR acquisition simulation. While it was expected to have no imaginary part in the cross-spectrum (because of the absence of ocean movement), a significant signal in the imaginary part was found suggesting that other phenomenon was responsible for some phase in the cross-spectrum. A joint theoretical and data analysis on both real and simulated data revealed that the small variation of the SAR observing geometry during the acquisition time has a significant contribution on the sub-look cross-spectra phase. This have an impact of the swell system signature and shall not be neglected when interpreting the cross-spectra for swell propagation ambiguity removal or, more generally, on phase exploitation strategies. A simple theoretical model has been developed to demonstrate the strong impact of this phenomenon on the cross-spectrum derivation and more specifically on its phase. We will show that this effect is linked to an important geophysical term that has been omitted in the classical literature leading to an incomplete SAR closed-form equation and, in turn to possible caveats in the operational wave inversion algorithm applied to Sentinel-1 products. In fact, considering this term is also critical to anticipate accurately future SAR missions performances 3 ESA Earth Explorer 10 - Harmony, understanding bistatic SAR The European Space Agency has selected Harmony as tenth Earth Explorer Mission. Harmony is a constellation of two companion satellites around Sentinel-1 platform. Among other capabilities, the two Harmony satellites will provide interferometric measurements and bi-static SAR acquisitions. This is a unique opportunity to better map ocean surface dynamics thanks to its enhanced azimuthal diversity. Based on the existing mono-static theoretical close-form, efforts have already been done by the Harmony scientific team to extend it to the bi-static case. However, the basic assumptions are still questionable and their validation with respect to numerical simulations highly desirable. The authors developed joint theoretical and numerical approaches to better understand the ocean wave mapping with bi-static SAR and started to explore the benefits of this configuration comparatively to the mono-static case. Preliminary studies show very encouraging results an reveal new possible strategies for mitigating SAR undesired effects such as azimuthal cut-off. 4 ESA SARWAVE project - sea state retrieval from Sentinel-1 large swath acquisitions The scientific teams involved in SARWAVE project (IfremerLOPS, Isardsat, TU Delft, ODL, DLR, CEOS-UP) proposed to develop and validate novel methods for sea state retrieval from Sentinel-1 Interferometric Wide Swath (IWS) products. This project, recently selected by ESA, started some month ago and recent results will be presented during a dedicated session of the Seasar 2023 workshop in Svalbard. Some of the methodologies presented in this abstract will be used in the algorithms development workpackage of the SARWAVE project. The latter is indeed an excellent opportunity to take benefit from the past and new research outputs dedicated to sea state retrieval from remote sensors (SAR, optical, altimeter, ...) and to exemplify the newly developed algorithm and produced sea-state L2 products over a one-year duration data-set of Sentinel-1 IWS acquisitions over European seas. References [1] G. Engen and H. Johnsen. Sar-ocean wave inversion using image cross spectra. IEEE transactions on geoscience and remote sensing, 33(4):1047–1056, 1995. [2] K. Hasselmann and S. Hasselmann. On the nonlinear mapping of an ocean wave spectrum into a synthetic aperture radar image spectrum and its inversion. Journal of Geophysical Research: Oceans, 96(C6):10713–10729, 1991. [3] H. E. Krogstad. A simple derivation of hasselmann’s nonlinear ocean-synthetic aperture radar transform. Journal of Geophysical Research: Oceans, 97(C2):2421–2425, 1992. On The Complementary Assimilation Of SAR And SWIM Wave Spectra In The CMEMS Global Wave System Lotfi Aouf1, Danièle Hauser2, Fabrice Collard3, Bertrand Chapron4 1Meteo France, France; 2LATMOS/CNRS; 3Ocean Data Lab; 4IFREMER corresponding author : [email protected] Directional wave observations from satellite missions play a crucial role to improve integrated sea state parameters, particularly swell dominant ones. The global CMEMS wave system is using in operations (Near Real Time) both SAR wave spectra from Sentinel-1 and SWIM wave spectra from CFOSAT. SAR instrument covers ranges of waves from roughly 200 m to 800 m of wavelength, while SWIM detects shorter scale of waves starting from 60 m to 500 m of wavelength. The objective of this work is to examine the combined assimilation of wave data from CFOSAT and Sentinel-1 missions. We investigated particularly the impact of assimilating directional wave spectra in unlimited fetch conditions such as in the Southern Ocean. Two years of operational wave products from MFWAM model will be used in the analysis and a control model run without assimilation will be performed. Particular attention will be considered to examine the underestimation of spectral wave energy from SAR for high wave height generated in severe storms in austral winter. Figure 1 shows the mean difference of Significant Wave Height (SWH) from model runs with assimilation of wave spectra from SAR of Sentinel-1 and SWIM of CFOSAT. This clearly indicates the underestimation of SWH from the assimilation of SAR wave spectra in Southern Ocean dominated by shorter long waves still wind dependent because of unlimited fetch conditions. In these conditions SWIM can capture correctly such wind-wave of wavelengths ranging between 50 to 150 m of wavelength. The overestimation of SWH from the assimilation of SAR spectra as illustrated by in reddish color in figure 1, reveals the limitation of SWIM to capture very long waves exceeding 500 m of wavelength, typically in west coast of central America and tropical ocean regions. Figure 1 : Mean difference of SWH from MFWAM model runs A and B for the period of May 2020. Run A is with the assimilation of SAR wave spectra, while run B is with the assimilation of SWIM wave spectra. The validation of model outputs has been performed with independent wave data from altimetry and also integrated sea state parameters provided by available drifting buoys. Figure 2 shows here the SWH bias maps in comparison with altimeters focused on Southern Ocean. We can easily see that the assimilation of SWIM wave spectra better reduces the bias in comparison with the assimilation of SAR wave spectra. (a) (b) (c) Figure 2 : bias maps of SWH from MFWAM model runs in comparison with altimeters (Jason-3, Saral and S3) for the period from May-August 2020. (a), (b) and (c) stand for runs without assimilation, with assimilation of SAR spectra only, and with assimilation of SWIM spectra only, respectively. The results show the relevance of using jointly SAR and SWIM wave spectra to take benefits from the best capturing of different scales of waves. The SWIM wave spectra can also be used to improve or calibrate SAR wave spectra in storm conditions of Southern ocean. We also examined the impact of using upgraded level 2 retrieval algorithms for CFOSAT on the wave forecast. In other respects we investigated the impact of using multi-mission directional wave observations on coupling parameters with ocean circulation model. Further results will be presented in the final paper. Unlocking the Potential of Distributed SwarmSAR for Highresolution Imaging of Ocean Waves M. M. M. Amir1, A. Bogoni1, P. L. Dekker2 1TeCIP Institute, Scuola Superiore Sant’Anna, Pisa, Italy email: {malikmuhammadharis.amir, antonella.bogoni}@santannapisa.it 2Delft University of Technology, Delft, The Netherlands email: [email protected] Abstract: The paper presents a novel concept of swarmSAR for improving azimuth resolution of decorrelating targets i.e., ocean surface in SAR imaging. SwarmSAR uses a swarm of small satellites to form a larger synthetic aperture, enabling higher resolution compared to traditional SAR. Results of the study show that swarmSAR offers a significant improvement in azimuth resolution compared to traditional SAR and provide insight into the limitations and future potentials. The paper suggests that this new technique has the potential to advance a range of applications, including environmental modelling, remote sensing, and oceanography. The study sheds light on the possibilities of using swarmSAR to tackle long-standing challenge in SAR imaging and promises to open up new avenues for research in this field. Introduction: Synthetic Aperture Radar (SAR) is a highly sophisticated remote sensing technology that plays a vital role in mapping the earth’s surface and obtaining detailed images. Its application spans across the fields, including geology, hydrology, meteorology, oceanography, and others [1]. SARs exploit the satellite movement to synthesize a long aperture antenna, observing the area of interest, and collecting radar echoes of the area while moving over it, to attain high azimuth resolution, which would otherwise need an antenna with an aperture of several kilometre. The ability of SAR to provide high-resolution images is a result of coherent integration of radar returns during the time of flight of a radar pulse along its path. However, the quality of these images can be impacted by various factors, including the aperture length. The amount of information collected by the radar is directly proportional to the aperture length, and hence, the aperture length is a critical factor affecting the resolution and accuracy of SAR images. The decorrelating nature of the target, because of the changes in ground scene during the flight duration can limit the aperture length. On land, the decorrelation is frequently not an issue because the ground is not changing during the period when SAR data is being gathered. Oceans, on the other hand, vary quickly over the duration of flight time, causing severe decorrelations, and limiting the permitted synthetic aperture length, and as a result affecting the azimuth resolution and SAR image quality [2]. To address these limitations, swarmSAR concept has emerged as a promising solution. The approach involves using multiple antennas in a close formation cooperating in a multiple-input multiple-output (MIMO) fashion. The idea behind SwarmSAR is to install several nodes, each of which has a basic imaging capacity that can be used independently. However, when they work together, they increase azimuth resolution and imaging capabilities [3]. In this paper, we address the swarmSAR concept for decorrelating target, to enhance the azimuth resolution. First, we will present the signal model, which will be followed by the discussion of preliminary results, and finally conclusion. The detailed results will be presented in the final version of the paper. Signal Model: Let us consider a decorrelating point target, with point target amplitude 𝐴 and a variable decorrelation amplitude 𝐴𝑑, having a correlation time 𝜏𝑐≈3.29𝜆𝑈 ⁄, where 𝑈 is the windspeed in this case we considered 𝑈 = 10𝑚𝑠 ⁄ [4]. The total amplitude of decorrelating point target 𝐴𝑡, which is product of 𝐴 and 𝐴𝑑. The decorrelating point target is illuminated by 𝑁 phase centres. Let Φ𝑖(𝑡,𝑅) represent the phase history of 𝑖𝑡ℎ satellite, where 𝑖 ={1,…,𝑁}, 𝑡 is the time and 𝑅 is the range. Φ𝑖(𝑡,𝑅)= 𝑒−2𝑘0𝑅𝑖 (1) Where, 𝑘0= 2𝜋 𝜆 ⁄, and 𝑅𝑖 is the range history of 𝑖𝑡ℎ phase centre. The integration length is set according to the correlation time, i.e., 𝑙𝑖≈𝜏𝑐.𝑣𝑠𝑐, where 𝑣𝑠𝑐 is satellite velocity. The monostatic received signal based on the phase history of the 𝑖𝑡ℎ phase centre can be written as: s𝑟,𝑖(Φ)= 𝐴𝑡.𝑒𝑗Φ𝑖 (2) Once the azimuth direction of the target is estimated, the signal from each satellite can be compressed in azimuth direction, this can be performed by using Fourier Transform (FT). S𝑟,𝑖(Φ)= ℱ{ s𝑟,𝑖(Φ)} (3) The compressed signals from different phase centres can then be combined and passed through the matched filter to generate azimuth compressed SAR image. 𝑆 = ∑𝑆𝑟,𝑖(Φ) 𝑁 𝑖=1 (4) Final SAR image can be expressed as; 𝐼 = ℱ−1(𝑆∗𝑀) (5) Here, 𝑀 is the matched filter and ℱ−1 represent inverse FT. Results: Before discussing the azimuth resolved images using monostatic and swarmSAR concept, let us discuss frequency response of signal from five satellites and simulation parameters, which are shown in figure 1 and table 1 respectively. Note that, since aperture length is same as the satellite separation, the frequency responses for different satellites is perfectly aligned together causing no overlapping or gaps between different apertures, and hence by combining all the apertures together, we are able to create a full long aperture, and hence increasing azimuth resolution. However, in a close to reality scenario, when the apertures are not perfectly aligned with each other, this will cause the frequency response to overlap which due to insufficient cut-off can lead to azimuth ambiguities. Figure 1 – Frequency Response for five satellites with the separation of 4 km between them Table 1 – Simulation Parameters Parameter Value Satellites Height 693 km Satellites Speed 7000 m/s Operating Frequency 2.9 GHz (S-Band) PRF 4000 Number of Satellites 5 Near Surface Ocean Wind Retrievals & Detection of Extremes Stoffelen, Ad; Ni, Weicheng; Xu, Xingou; Portabella, Marcos; Makarova, Evgeniia; Cossu, Federico; Sánchez Rabaneda, Alberto Extreme Winds From SAR And Their Use For KuAnd C-Band Wind Scatterometer Resolution Enhancement Zecchetto, Stefano; Zanchetta, Andrea; Sclavo, Mauro; Shamsaddini, Parsa; Keshavarz, Ahmad High-Resolution SAR Winds from Deep Learning in Coastal Areas Khan, Salman; Young, Ian; Ribal, Agustinus; Canto, Marites; Davy, Robert; Hemer, Mark Australian Coastal SAR Ocean Winds: Data, Portal, and Next products Dimitriadou, Krystallia; Badger, Merete; Hasager, Charlotte Bay; Olsen, Bjarke Tobias SAR for Offshore Wind Fields in the Mediterranean Sea Hindberg, Heidi; Espeseth, Martine; Johnsen, Harald; Tollinger, Mathias Operational Wind Retrieval Using Cross-Polarization And Doppler Data Marquart, Robin; Mouche, Alexis; Chapron, Bertrand Analysis Of SAR Ocean Scenes Texture For Wind Direction Retrieval And Generation Of Synthetic Images EXTREME WINDS FROM SAR AND THEIR USE FOR KU-BAND AND C-BAND WIND SCATTEROMETER RESOLUTION ENHANCEMENT Ad Stoffelen1*, Weicheng Ni2, Xingou Xu3, Marcos Portabella4, Evgeniia Makarova4, Federico Cossu4 and Alberto Rabaneda5 1. Royal Netherlands Meteorological Institute KNMI, 3730 AE,De Bilt, Utrecht, The Netherlands 2. National University of Defense Technology,Changsha 410073, China 3. Key Laboratory of Microwave Remote Sensing, National Space Science Center, CAS, Beijing, 100190, China 4. Barcelona Expert Centre (BEC), Institut de Ciències del Mar (ICM-CSIC) Barcelona, Spain 5. Norwegian Meteorological Institute, Oslo, Norway * [email protected] Introduction Tropical cyclones (TC), or hurricanes cause severe damage during their land-falls, in China alone, the average 9.3 land-fall TCs cause more than ＄1 billion property loss while they take thousands of human lives each year. At the same time, remotely sensed data have the potential of providing vital information for improved forecasts. Among them, Synthetic Aperture Radar (SAR) images capture the most dynamic detail. European, Canadian and Chinese efforts to capture TCs have resulted in a considerable data base of a few hundred TC SAR acquisitions. Exploiting these data, we report on SAR image classification of TC features [1], TC wind direction retrieval using features in all polarizations [2,3], intercalibration of ECMWF, SAR and ASCAT scatterometer winds in a triple collocation [4]. While SARs provide detail, wind scatterometers provide the capability to track TCs when exploiting the extending scatterometer virtual constellation. Scatterometers have been providing high-quality ocean surface wind products for more than 40 years [11], while they provide good extreme-wind retrieval on their native resolution of about 25 km [5]. Ku-band scatterometers have in principle similar capability, but suffer from heavy rain contamination and where quality control and rain correction are in progress [6,7]. Recently, with reference to the Step-Frequency-MultichannelRadiometer (SFMR) products [8], the quality of the C-band extreme wind speeds have been well proven and extension to Ku-band scatterometers is ongoing, tackling the problem of contamination by precipitation, as further illustrated below. Furthermore, improved extreme wind vector retrieval is being explored for C-band and later Ku-band scatterometers by resolution enhancement with reference to the extended SAR data base mentioned above using parametric high-resolution TC models [9]. Since sufficient SAR and scatterometer collocations for statistical enhancement techniques for varying hurricane categories are now available, the standard 2DVAR processing for scatterometer ambiguity removal [10] is employed for resolution enhancement of scatterometer winds in TCs, as illustrated in next secction. Work in Progress In-situ wind speeds are inconsistent for winds higher than 15 m/s, which poses a problem for wind speed calibration in TCs. The highest quality regular winds are from moored buoys and these are available up to about 25 m/s in sufficient statistical quantity for analysis [12]. Dropsondes from airplane campaigns by the NOAA hurricane hunters [13], which generally available above 20 m/s and are used to calibrate the Step-Frequency Microwave Radiometer (SFMR) winds. SFMR winds may subsequently be used for satellite wind calibration, due to their abundance. At 25 m/s there is an approximate 40% difference in wind strength between wind calibrations based on SFMR/dropsondes and moored buoys, affecting all satellite wind retrievals and physical modelling attempts of exchange processes at extreme winds. A more detailed assessment of dropsonde characteristics is needed [12], hence probably also affecting all SAR wind products [4]. For reconciliated speeds between satellite instruments [4], we attempt spatial resolution enhancement and tracking of TCs or polar lows using the virtual constellation of microwave sensors, both active and passive. The ASCAT C-band scatterometer TC acquisitions are now being tested as depicted in Figure 1. Figure 1. Illustration of ASCAT resolution enhancement: (a) TC TEDDY (acquired on September 16, 2020) imaged by ASCAT VV winds; (b) ECMWF forecast TC TEDDY winds resampled onto ASCAT; (c) 2DVAR wind analysis increments corresponding to (a) and (b). An apparent TC eyewall is generated around the TC center; (d) Super resolution ASCAT result; (e) TC TEDDY imaged by collocated SAR winds. In the 2DVAR analysis the typical spatial errors in the ECMWF forecasts need to be characterized in terms of longitudinal and transverse wind components. These spatial errors are particular to the hurricane under analysis and depend on the hurricane intensity and Radius of Maximum Wind (RMW), which can be estimated from the ASCAT inputs. The next step is to extend the resolution enhancement work to Ku-band scatterometers, thereby profiting from ongoing advancements in wind calibration, quality control and in particular rain screening and possibly rain correction for Ku-band scatterometers [6,7]. Conclusions The extensive SAR campaigns on acquisitions of tropical cyclones have resulted in a data base that allows the innovation of SAR wind products for TCs, providing unprecedented detail on the surface winds, supplementing detailed wind information, otherwise only available from hurricane campaigns. The many SAR TC acquisitions provide a good statistical basis for enhancing TC winds from the extending virtual wind scatterometer constellation, allowing improved tracking of TCs over their lifetime. Some of these enhancement techniques for surface winds may also be employed for microwave radiometers winds in future work. References [1] Weicheng Ni, Ad Stoffelen & Kaijun Ren (2022), Hurricane eye morphology extraction from SAR images by texture analysis. Front. Earth Sci. 16, 190–205. https://doi.org/10.1007/s11707-021-0886-9. [2] Weicheng Ni, Ad Stoffelen & Kaijun Ren (2023),Tropical Cyclone Wind Direction Retrieval From Dual-Polarized SAR Imagery Using Histogram of Oriented Gradients and Hann Window Function", IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 878-888, https://doi.org/10.1109/JSTARS.2022.3230441. [3] Weicheng Ni, Ad Stoffelen & Kaijun Ren (2022),"Tropical Cyclone Wind Direction Retrieval Using Histogram of Oriented Gradients on Dual-Polarized Synthetic Aperture Radar Images", IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium, Kuala Lumpur, Malaysia, pp. 6967-6970, https://doi.org/10.1109/IGARSS46834.2022.9884396. [4] Weicheng Ni, Ad Stoffelen, Kaijun Ren, Xiaofeng Yang, and Jur Vogelzang (2022), "SAR and ASCAT Tropical Cyclone Wind Speed Reconciliation" Remote Sensing 14, no. 21: 5535. https://doi.org/10.3390/rs14215535. [5] Federica Polverari et al. (2022), "On High and Extreme Wind Calibration Using ASCAT", IEEE Transactions on Geoscience and Remote Sensing 60, 1-10, Art no. 4202210, https://doi.org/10.1109/TGRS.2021.3079898. [6] Xingou Xu, Ad Stoffelen, Marcos Portabella, Wenmig Lin and Xiaolong Dong (2021), "A Comparison of Quality Indicators for Ku-Band Wind Scatterometry & for Typhoons Lekima and Krosa", 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, Brussels, Belgium, 7584-7587, https://doi.org/10.1109/IGARSS47720.2021.9553678. [7] Ke Zhao, Ad Stoffelen, Jeroen Verspeek, Anton Verhoef and Chaofang Zhao (2023), Bayesian algorithm for rain detection in Ku-band scatterometer data, IEEE Transactions on Geoscience and Remote Sensing, in revision. [8] F. Polverari, J. W. Sapp, M. Portabella, A. Stoffelen, Z. Jelenak and P. S. Chang (2022), "On Dropsonde Surface-Adjusted Winds and Their Use for the Stepped Frequency Microwave Radiometer Wind Speed Calibration", IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-8, Art no. 4208308, https://doi.org/10.1109/TGRS.2022.3189310. [9] W. Ni, A. Stoffelen, K. Ren and X. Yang, "Tropical Cyclone Intensity Estimation From Spaceborne Microwave Scatterometry and Parametric Wind Models", IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 15, pp. 4719-4729, 2022, https://doi.org/10.1109/JSTARS.2022.3180281. [10] Jur Vogelzang, Ad Stoffelen (2018), Improvements in Ku-band scatterometer wind ambiguity removal using ASCATbased empirical background error correlations. Q J R Meteorol Soc. 144: 2245– 2259. https://doi.org/10.1002/qj.3349. [11] Jur Vogelzang & Ad Stoffelen (2021). Quadruple collocation analysis of in-situ, scatterometer, and NWP winds. Journal of Geophysical Research: Oceans, 126, e2021JC017189. https://doi.org/10.1029/2021JC017189. [12] Ad Stoffelen et al., "Hurricane Ocean Wind Speeds", 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, Brussels, Belgium, 2021, pp. 1182-1185, https://doi.org/10.1109/IGARSS47720.2021.9554667. [13] https://www.aoml.noaa.gov/hurricane-research-division/. Analysis of SAR ocean scenes texture for wind direction retrieval and generation of synthetic images Robin Marquart, Alexis Mouche, Bertrand Chapron January 2023 Abstract This paper provides an extensive analysis of high resolution C-band ocean SAR scenes. Now routinely provided by Sentinel-1 acquisitions at global scale in wide swath and wave modes, C-band detected ocean surface roughness changes generally trace local air-sea marine-atmosphere layer conditions. Building on these large data sets, a brute force approach is first applied to establish a robust relationship between the local radar texture and the wind direction. Training a convolutional neural network model, performances can be assessed against co-located WindSat radiometer estimates, with respect to the radar configuration (polarization, incidence angle, antenna-towind direction), the wind speed and the nature of the air-sea interactions. In presence of atmospheric rolls, the key textural information lies between 800 and 1600 m. Moreover, to interpret and then synthesize a realistic scene, the texture spectral information is solely necessary. This does not apply for scenes where convective cells dominate the texture. For such cases, the detected scale organisation matters. To interpret and synthesize a realistic scene, both spectral and phase distributions must be considered. The present analysis thus provides new means to derive improved estimates of very local air-sea marine-atmosphere layer conditions. 1 Introduction It has long been recognized that high-resolution SAR ocean scenes quite systematically exhibit atmosphere signatures. For scales larger than 300m, filtering out ocean swell contributions and slick impacts, most radar detected roughness changes characterize the nature of the stability above the sea surface. Several studies thus already largely investigated the relationship between these amplitude modulations with stability to possibly retrieve the MoninObhukov length from SAR large scenes [4] or to characterize the surface layer stratification regime associated to particular patterns [1]. Based on Sentinel-1 global Wave Mode data, these authors evidenced that signatures of atmospheric rolls were prevalent for near-neutral conditions while signatures of convective cells were appearing for unstable conditions. Recently, methodologies based on convolutional neural networks (CNN) has been used to provide efficient multi-scale decomposition of instantaneous SAR ocean image textures. In particular, [3] applied such an approach to classify wave mode images into 10 different geophysical classes. An other study [5] also relies on such a decomposition to evidence links with the local wind direction estimates. In that context, CNN methods largely extend Fourier-based methodology, mostly useful to estimate the local direction of atmospheric rolls. In the present study, the objective is to primarily follow and extend the approach presented in [5] to a broader frame, including all sea conditions and Sentinel-1 modes. From the extended analysis, a more precise investigation of the small-scale organisation, spectral and phase distributions, are then performed. The proposed methodology is especially useful to better understand the links between SAR texture statistical properties, the surface wind direction and local air-sea marine-atmosphere layer conditions. 2 Methodology 2.1 Dataset and neural network A dataset was built for each of the following Sentinel-1 exclusive acquisition modes: Extra-Wide swath (EW), Interferometric swath (IW), Wave (WV). These are three collections of collocated patches extracted from SAR acquisitions and a reference wind direction. Each patch is defined by a size and a resolution. As a first choice, we set the size to be around 18 km, corresponding to the size of an image acquired in WV. A pixel resolution of 400 m is then considered to filter most of the swell signature in the signal. The reference wind direction is given by 1 Figure 1: (a) A streaks scene as fed to the CNNs. (b) Previous streaks scene with randomized phase. (c) Artificially generated streaks. satellite passive microwave WindSat estimates. A pair is considered to be robust when the WindSat wind direction estimates is consistent (±10◦) to the one given by the ERA-5. WindSat observations may greatly deteriorate in presence of heavy precipitations [2]. Only cases with a zero rain rate are kept. Note, heavy rain signatures on SAR images can hide features related to the wind direction to possibly prevent any learning from the neural network. For each of these datasets, several deep residual networks (ResNet) are trained. As each model is initialized randomly and trained independently, this allows to predict an ensemble of different wind direction values and then determine which direction is the most probable for a given patch as well as the associated uncertainty. The patches resolution has been degraded to the same 400 m resolution for the three acquisition modes, the size chosen to fit in the wavemode swath, to take benefit of the transfer learning method. The models are first trained on patches extracted from WV acquisitions for which we have a much larger dataset. The WV model coefficients are then transferred on IW and finally on EW for which less data is available. 2.2 Synthetic images In this study we also attempt to create synthetic SAR scenes. Two methods have been tested : the modification of an existing image and the creation from scratch. This is done in order to help discussing the results given by the neural networks, further analyze the signature of phenomena of interest and to allow for data augmentation or multi-resolution analysis. Here is presented how those artificial scenes are synthesized. To begin with, starting from patches used in the validation datasets, we removed the part of the signal related to the organization of the phases in the Fourier domain : we only kept the modulus of the spectrum and randomized the phase component. This is done by replacing the phase component of the considered patch by the one of a Gaussian white noise. We also directly synthesized streaks scenes by modulating white noise with a correlation function defined such as the noise level, width, length, regularity and orientation of the streaks. These parameters can then be can be tweaked to test different configurations. An example is given on figure 1. Finally, to examine the importance of the scales studied, we also applied low pass and high pass filters to the patches of interest at different lengths. 3 Results & Discussion The neural networks trained on each acquisition mode provide reliable estimations of the wind direction. In average, the Mean Absolute Error (MAE) between the prediction and the reference value is inferior to 12◦with a standard deviation (STD) below 19◦for IW and WV. For EW the MAE is lower than 15◦with a STD below 24◦. The bias is lower than 2◦across all acquisition modes. The best wind estimations are obtained for wind speeds between 10 and 20 m/s. In this specific range, the MAE drops below 7◦for IW, 9◦for WV and 11◦for EW. Moreover, the predictions do not seem to be dependent of the incidence angle in the case of WV or IW where plenty of data is available across the whole incidence range. In EW, the lower amount of data available in the training dataset at incidence inferior to 27◦does seem to have an impact on the performances. The error linearly decreases from 20◦ to 15◦between incidence angles of 20◦and 27◦. In the contrary, the quality of inference is significantly impacted by the nature of the atmospheric phenomena visible in a given patch. The performances of the model with respect to 10 different geophysical classes as defined in [3] for WV has been assessed. The best results were obtained on streaks events where the mean difference between predictions and reference is lower to 9◦for both WV1 and WV2. On micro convective cells or rain cells situations, this difference increases by more than 4◦. An illustration of those results is given on figure 2. Even tough the model was trained only on the VV polarization channel, it is still possible to infer on the VH polarization channel. In fact, in the case of extreme waves events, the swell signature can hinder the wind direction 2 Figure 2: Example of the results obtained on different type of scenes. (a) (b) Figure 3: Sentinel-1 acquisitions of the tropical cyclone Surigae with windfield estimation from the neural networks. (a) Polarization VV. (b) Polarization VH retrieval because its unusually large wavelength can be mixed up with rolls by the neural models. This is expected to impact more importantly the VV than the VH polarization channel as the wave modulation is higher in VV than in VH due to the tilting effect. Considering the VH channel is thus interesting in these cases. The figure 3 illustrates this aspect on a Sentinel-1 acquisition over the Surigae cyclone where the inference made on the VH channel seems more realistic (wind swirling around the eye on the right hand side of TC tracks where swell are expected to be the longest) than the one from the VV channel. The models were also used to predict the wind direction on synthetic patches. In the case of wind streaks, when the phases of a given patch are randomized in the Fourier domain, the models are still able to perform a reliable estimation. However, this does not apply in the case of micro convective cells. This means that the organisation of the cells-induced structure is essential to retrieve the wind direction. Progressively filtering different levels of scales using low pass and high pass filters allowed to estimate what are the most important ones for the wind direction retrieval. It appears that it is between 800 m and 1600 m that lies the majority of features directly related to wind direction estimation. References [1] Justin E. Stopa, Chen Wang, Doug Vandemark, Ralph Foster, Alexis Mouche, and Bertrand Chapron. Automated global classification of surface layer stratification using high-resolution sea surface roughness measurements by satellite synthetic aperture radar. Geophysical Research Letters, 49(12). [2] Meissner Thomas and Frank Wentz. Wind-vector retrievals under rain with passive satellite microwave radiometers. Geoscience and Remote Sensing, IEEE Transactions on, 47:3065 – 3083, 10 2009. [3] Chen Wang, A. Mouche, P. Tandeo, J. Stopa, B. Chapron, R. Foster, and D. Vandemark. Automated geophysical classification of sentinel-l wave mode sar images through deep-learning. IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, pages 1776–1779, 2018. [4] George S Young, Todd D Sikora, and Nathaniel S Winstead. Inferring marine atmospheric boundary layer properties from spectral characteristics of satellite-borne sar imagery. Monthly weather review, 128(5):1506– 1520, 2000. [5] Andrea Zanchetta and Stefano Zecchetto. Wind direction retrieval from sentinel-1 sar images using resnet. Remote Sensing of Environment, 253:112178, 2021. 3 HIGH-RESOLUTION SAR WINDS FROM DEEP LEARNING IN ARCTIC FJORDS Stefano Zecchetto1,2, Andrea Zanchetta3, Mauro Sclavo1, Ahmad Keshavarz2, and Parsa Shamsaddini2 1Istituto di Scienze Polari, National Research Council of Italy, Padova, Italy 2Faculty of Intelligent Systems Engineering and Data Science, Persian Gulf University, Bushehr, Iran 3Department of Music, Hong Kong Baptist University, Hong Kong SAR, China Abstract A deep learning methodology based on a residual neural network (ResNet), developed to retrieve wind directions from SAR at 500 m grid without external information, opens the possibility to investigate the spatial characteristics at mesoscale γin small semi-enclosed areas, like lagoons and Arctic fjords. Avoiding to illustrate technical details concerning the ResNet methodology, this contribution is focused on showing a ResNet wind field over Svalbard fjords. The richness of details provided makes the comparison with other data sets (models/in-situ) somehow insufficient, simply because they do not have a resolution similar to that of SAR wind. We tackle this issue by adding other approaches to the standard co-located comparisons, like the analysis of the spatial gradient of wind. In the example reported in this work and in literature it turns out that the structure of the ResNet wind fields is compatible with some of the meso-scale γfeatures of the wind. Key words: Deep residual network, Arctic Fjords, Synthetic Aperture Radar (SAR), Wind field. 1. INTRODUCTION Retrieval of wind field from SAR needs, among others, the knowledge of the wind direction. The main techniques at present available are reviewed in Zecchetto and Zanchetta (2022), to which the interested reader is addressed. Summarizing by considering only the methods evaluating wind direction directly from the SAR images, we pointed out that the intrinsic limitations of the existing methodologies (requiring the presence of wind-induced streaks or weakly divergent fields or working only in deep water open sea) claim for a robust methodology able to derive high resolution (below 1 km) wind fields both in open sea and coastal areas. Thus, a deep learning methodology based on the residual convolutional neural network, hereafter referred as ResNet, has been developed using Sentinel-1 SAR images (Zanchetta and Zecchetto, 2021) and tested over different meteorological situations and areas (Zecchetto and Zanchetta, 2022). From a statistical point of view, the ResNet wind directions agree very well with those from ECMWF and in-situ data; from the applicative point of view, the 500 m spatial grid of ResNet SAR winds attains an exhaustive coverage of small enclosed sea areas, like the Venice Lagoon, allowing the investigation of the inner structure of the wind fields up to the meso-scale γ(Orlanski, 1975), i.e. below 10 km. In our knowledge, there is not any other methodology to derive the wind from SAR with such richness of detail and without any external information. 2. THE DATA The C-band Sentinel-1 images IW GRDH L1 at VV polarization and pixel size of 10 meters (European Space Agency, 2013), downloaded from the ESA Sentinels Scientific Data Hub1, have been used in this work. In-situ and model data have been used here as reference: they are the in-situ data in the western coast of Svalbard, downloaded from the Norwegian Meteorological Institute2, and the AROME Arctic 2.5 km model wind data are the forecasts at 1 hour sampling by the Norwegian Meteorological Institute. 1https://scihub.copernicus.eu/ 2frost.met.no/api.html 2 3. IMAGE PRE-PROCESSING AND WIND DIRECTION COMPUTATION BY RESNET The original SAR images underwent to a pre-processing, carried out using SNAP3free software, i.e. precise orbit determination, thermal noise reduction, calibration, de-speckle filtering, terrain correction, land-sea mask and down-sampling to 50 m. ResNet works from these pre-processed images on partitioned, partly overlapping, sub-images of 64 × 64 pixels (3.2 km × 3.2 km), centered on a regular grid of 500 m, previously filtered with a Gauss filter to remove the ocean wave signatures. Therefore, the vectors closest to the land are at least 1.5 km offshore. ResNet has been trained with 806139 samples derived from 25 Sentinel-1 images, in a similar way as in Zanchetta and Zecchetto (2021). 4. RESULTS Figure 1 reports a Sentinel-1 image of the southwestern part of Svalbard (left panel), along with the ResNet wind field (right panel). The image was taken in wintertime, and sea-ice is present both as pancake and first year ice. As the wind cannot be derived over ice, we developed a methodology (Shamsaddini et al., 2023), based on the Gray Level Cooccurrence Matrix, to obtain sea-ice masks, contoured by white lines in the left panel. The image shows a tongue of high backscatter in the Isfjorden area, (between 9.5◦and 17.5◦E and 77.8◦and 78.4◦N), well visible in the ResNet wind (left panel), showing an impressive number of details: the down-fjord wind in Isfjorden, the katabatic flows west of Prins Karls long island (≈10◦E, 78◦N). The bias between AROME model and SAR winds shows a good agreement (0.4 m/s and 3◦), but cannot show whether the several short scale variations provided by SAR are artifact of the method used or indeed represent meso-scale γcharacteristics of the wind. For this reason, we investigate the wind speed and direction variability of the down-fjord wind in the Isfjorden shown in the right panel of Figure 1. Figure 2 reports the wind speed Figure 1. Example of SAR wind field over the Svalbard. Left panel: Sentinel-1A of 2 February 2020, 06:15 UTC. The white contours identify the presence of ice. Right panel: ResNet wind field in the sea areas (sea-ice areas in white). Black arrows are in-situ winds (not in scale). The black dashed line is the transect along which the wind variability is studied. and direction variability along the transect from ResNet, AROME and ECMWF model data. Apart from the obvious difference in details between ResNet and model winds because of the different grid size, the models variations differ in both wind speed and direction, with some space lag with variations of the SAR winds. This shows the issues arising when comparing SAR winds with models in small areas. The spatial variability of the wind speed and direction are of 0.8 (m/s)/km and 8◦/km, with a dominant wavelength of 7.1 km, numbers close to those obtained from the analysis of high frequency in-situ and SAR winds in the northern Adriatic sea (Zecchetto and Zanchetta, 2022) (not possible with the 1 h in-situ data available here). This indirectly confirms the reliability of the estimates provided by ResNet. 3https://step.esa.int/main/toolboxes/snap/ 3 5. FINAL CONSIDERATIONS The ResNet SAR wind fields describe, in unprecedented detail, the spatial characteristics of the wind even in very demanding areas as the Arctic fjords. The richness of details provided makes somehow insufficient the comparison with other data set, simply because the present available data set (models/in-situ) do not have resolutions comparable to that provided by ResNet. Thus, we have to envisage other approaches, different from the standard co-located data comparison, approaches testing the compatibility rather than bias of RMS difference. One of these is the analysis of the wind spatial gradient that has shown, in the example reported here as well as in Zecchetto and Zanchetta (2022), that the spatial structure of the ResNet wind fields is compatible with some features of the wind in the meso-scale γrange. An other approach could be the comparison with other methods deriving the wind from SAR without external information, such as the 2D Continuous Wavelet Transform (Zecchetto, 2018). Figure 2. Variability of wind speed (left panel) and direction (right panel) along the transect depicted in the right panel of Fig. 1. ACKNOWLEDGMENTS The Sentinel-1 images have been downloaded from the ESA Sentinels Scientific Data Hub. The in-situ data have been obtained from the Norwegian Meteorological Institute web site www.met.no. The work has been partially funded by the Italian Space Agency (ASI), Contract Call ASI DC-UOT-2019-017, Project APPLICAVEMARS n. 2021-4-U.0 (CUP F65F21000070005). REFERENCES European Space Agency (2013). Sentinel-1 User Handbook. ESA Standard Document. Available online at https://sentinel.esa.int/. Orlanski, I. (1975). A rational subdivision of scales for atmospheric processes. Bull. Amer. Meteor. Soc., 56:527–530. Shamsaddini, P., Keshavaz, A., and Zecchetto, S. (2023). Wind-robust sea-ice discrimination from sentinel-1 texture features. In ICEE 2023, volume submitted. Zanchetta, A. and Zecchetto, S. (2021). Wind direction retrieval from Sentinel-1 SAR images using ResNet. Remote Sensing of Environment, 253. Zecchetto, S. (2018). Wind Direction Extraction from SAR in Coastal Areas. Remote Sensing, 10(2):261. Zecchetto, S. and Zanchetta, A. (2022). Structure of high resolution SAR winds over the Venice lagoon area. IEEE Trans. Geos. and Remote Sensing. resolution is 10 min. The median wind speed for every 10 minutes is used, indicating that wind gusts are ignored. Figure 1 below shows the wind direction and speed from the platform, where the majority of the wind comes from southwest, west, and north-west. The average wind is 7.6 m/s, the middle wind is 7.2, and the maximum wind is 27 m/s throughout the year 2022. Fig. 1 In-situ wind direction and speed from NORNE platform during 2022. III. INVERSION METHOD SAR wind retrieval is typically done by defining a cost function which is a function of the true wind vector (speed and direction) and depends on the measured SAR backscatter and a prior wind vector as an input. The measured SAR backscatter is related to the wind vector using a forward model. Working from [6], the full cost function is defined as Here, we use the CMOD5.n [7] model for co-polarization backscatter 𝜎𝑚𝑜𝑑 𝑉𝑉 , the MS1A [8] model for cross-polarization backscatter 𝜎𝑚𝑜𝑑 𝑉𝐻 , and the CDOP [4] model for the DCA 𝑑𝑐𝑎𝑚𝑜𝑑 𝑉. The true wind vector is described by its speed 𝑤 and its direction 𝜙, and the prior wind direction 𝜙𝑝𝑟𝑖𝑜𝑟 is the ERA5 ECMWF [9] model wind direction. The model wind speed is not used directly in the cost function; however, the model wind speed and direction are both used as starting points in the inversion procedure. The selection of the weights in the denominators will assign the relative importance of matching this particular term in the cost function. For instance, with a high 𝑠𝜙 it will be less important in the inversion procedure to match 𝜙 with 𝜙𝑝𝑟𝑖𝑜𝑟. In this case we select 𝑠𝑉𝑉 = 𝑠𝑉𝐻 = 0.1, 𝑠𝑑𝑐𝑎 = 0.4 and 𝑠𝜙= 5. Over ocean the cross-polarization backscatter will hit the noise floor for low wind cases, and the co-polarization backscatter will saturate for high wind cases. Ideally, we should use the cross-pol signal for high wind and the co-pol signal for low wind. This is regulated through the 𝑊(𝜎𝑜𝑏𝑠 𝑉𝐻 ) term in (1). We select a wind speed interval [𝑤𝑙, 𝑤ℎ], where for all wind speeds below the interval we only use co-pol and for wind speed above the interval we only use cross-pol. We use an interval of [8,30] m/s and define the weight function as Wind inversion is performed with four combinations of input data and thus four different cost functions: - based on co-polarization data alone (copol), using terms 1 and 2 in (1). - based on a combination of co-polarization and cross-polarization (dual), using terms 1, 2, and 3 in (1). - based on the DCA and coand cross-polarization data (full), using all terms in (1). - based on co-polarization data and the DCA (codop), using terms 1, 2 and 4 in (1). The copol and dual scheme is applied to both S1 and RS2 datasets, and the results are processed from the standard GRD products using the CryoniteOcean software running at KSAT. The full and codop is limited to the S1 dataset as the DCA is not available for the Radarsat-2 dataset, and we perform the inversion starting from the S1 level 2 OCN products. IV. PRELIMINARY RESULTS Figure 2 shows the results for the four different inversion methods for a S1 scene on 20220508. Here we see that the copol, dual and full results for wind speed are quite similar, while the codop has some higher winds at near range. Either way the SAR wind speed is much more detailed than the model wind speed. The DCA is estimated at a resolution of 3x3 km, while the SAR wind has a resolution of 1x1 km, we can see this difference in resolution in the level of details comparing codop/full with copol/dual. For the direction, as expected the copol direction is very close to the model direction. The dual direction has more variations. We can also see the 3 swath seams in the dual direction, this is due to the inferior noise correction applied on the cross-pol backscatter. Fig. 2 Results for S1A IW GRDH 1SDV 20220528T055454. Upper row is wind speed and lower is wind direction, using the four different inversions and the ECMWF model data. Figure 3 shows screenshots from an interactive dashboard with statistics of the in-situ wind speed and direction from NORNE versus the four methods (copol, dual, codop, and full). The boxes below the scatter plots show the standard deviations of SAR wind and SAR wind directions versus in-situ. Codop and full-pol deviates the most from the in-situ wind data compared to the other methods, while the copol and dual-pol show similar performance. Preliminary results indicate that the use of crosspol is useful for increasing the information about the wind direction. Visually the retrieved wind speed seems deteriorated because of the swath seams, showing the need for more robust noise corrections. The use of DCA deteriorates the results, the weighting needs to be investigated further. Fig. 3 Upper: Time series of SAR wind and direction from in-situ, copol, codop, and dual-pol. Lower: Scatterplots of SAR winds versus insitu wind from copol, dual, co-dop, and full, respectively. REFERENCES [1] Horstmann, J.; Schiller, H.; Schulz-Stellenfleth, J.; Lehner, S. “Global wind speed retrieval from SAR”. IEEE Trans. Geosci. Remote. Sens. 2003, 41, 2277–2286. [2] Vachon, P.W.; Wolfe, J. “C-band cross-polarization wind speed retrieval”. IEEE Geosci. Remote. Sens. Lett. 2010, 8, 456–459. [3] Zhang, B.; Perrie, W. “Cross-polarized synthetic aperture radar: A new potential measurement technique for hurricanes”. Bull. Am.Meteorol. Soc. 2012, 93, 531–541. [4] Mouche, A.A.; Collard, F.; Chapron, B.; Dagestad, K.F.; Guitton, G.; Johannessen, J.A.; Kerbaol, V.; Hansen, M.W. “On the use of Doppler shift for sea surface wind retrieval from SAR”. IEEE Trans. Geosci. Remote Sens. 2012, 50, 2901–2909. [5] S-1 Mission Performance Centre, “Thermal Denoising of Products Generated by the S-1 IPF”, https://sentinel.esa.int/documents/247904/2142675/ThermalDenoising-of-Products-Generated-by-Sentinel-1-IPF [6] Tollinger M.; Graversen R.; Johnsen H. "High-Resolution Polar Low Winds Obtained from Unsupervised SAR Wind Retrieval", Remote Sens. 2021, 13(22), 4655; https://doi.org/10.3390/rs13224655 [7] Hersbach H., “CMOD5.N: A C-band geophysical model function for equivalent neutral wind,” ECMWF, Reading, U.K., 2008. Tech. Memo. 554. [8] Mouche, A.A.; Chapron, B.; Zhang, B.; Husson, R. “Combined co-and cross-polarized SAR measurements under extreme wind conditions”. IEEE Trans. Geosci. Remote Sens. 2017, 55, 6746–6755. [9] Hersbach, H. et al., “The ERA5 Global Reanalysis” May 2020 Doppler Shift Retrievals Doppler Shift Retrievals Kleinherenbrink, Marcel; Yuan, Yan; Theodosiou, Andreas; Gaultier, Lucile; Collard, Fabrice; Chapron, Bertrand; Lopez-Dekker, Paco Multiscale Effects on Harmony's High-resolution Ocean Observations Elyouncha, Anis; Eriksson, Leif; Gommenginger, Christine Observations of the Agulhas Current by Along-track Interferometric Synthetic Aperture Radar Romeiser, Roland Review of TerraSAR-X Based Current Retrieval Activities at the University of Miami Martin, Adrien; Macedo, Karlus; McCann, David; Portabella, Marcos; Marié, Louis; Marquez, José; Carrasco, Ruben; Duarte, Rui; Meta, Adriano; Gommenginger, Christine; Martin-Iglesias, Petronilo; Casal, Tania OSCAR: A New Airborne Instrument To Image Ocean-Atmosphere Dynamics At The SubMesoscale Moiseev, Artem; Collard, Fabrice; Johannessen, Johnny A Ocean Surface Currents from Sentinel-1 Doppler observations Guitton, Gilles; Collard, Fabrice; Johnsen, Harald; Engen, Geir; Recchia, Andrea; Cotrufo, Alessandro; Bras, Sergio; Miranda, Nuno; Pinheiro, Muriel Towards Calibrated Sentinel-1 OCN RVL Products Domps, Baptiste; Guérin, Charles-Antoine Evaluation of Surface Currents Derived from Sentinel-1 SAR Doppler Shift in the Northwestern Mediterranean Sea Using Coastal HF Radars Multiscale effects on Harmony’s High-resolution Ocean Observations M. Kleinherenbrink, Y. Yuan, A. Theodosiou, L. Gaultier, F. Collard, B. Chapron, P. Lopez-Dekker January 31, 2023 1 Introduction The two satellites of the Harmony mission will fly in constellation with a Sentinel-1 Synthetic Aperture Radar (SAR) satellite. Harmony’s passive radar instruments will receive signals transmitted by Sentinel-1 after reflection from the surface. The multistatic system uses three lines-of-sight to observe high-resolution normalized radar cross section (NRCS) and Doppler. With Harmony’s high-resolution observations we aim to observed high-resolution O(1 km2) wind-stress anomalies and ocean-surface currents. The interpretation of high-resolution signals is not trivial as they are affected by non-locally generated waves. Long waves O(>5 m) have relaxation scales much larger than the intended resolution of the Harmony level-2 products. The long waves affect the NRCS and the Doppler in different ways [6, 7, 5, 3]. Estimation of wind stress and ocean-surface current should therefore not be done on a pixel basis, but it requires multi-scale dynamical approaches. In this paper, we show the result of forward modelled polarimetric NRCS and Doppler for Harmony and Sentinel-1. Input for our models are wind (stress) and ocean current grids from models. We use an equilibrium short-wave spectrum with a parametric long-wave spectrum to estimate the NRCS and Doppler. Alternatively, we use the Simulating WAves Nearshore (SWAN) model[1] for the long waves to capture nonlocal effects of wind and currents. By comparing both, we demonstrate why a pixel-based inversion is not suitable for highresolution products. 2 Methodology The modelling of NRCS and Doppler is based on a bistatic implementation of the Radar Imaging Model (RIM)[6, 7] and the Doppler RIM (DopRIM)[5]. The forward models require as input a wave spectrum at each grid cell that is computed from the wind and ocean currents in the scene. 2.1 Wave spectra The wave spectrum is split at ten times the peak wavenumber, kl= 10kp, into a shortand a long-wave spectrum. The twodimensional curvature spectrum depends on the wavenumber vector  kand is described by B( k) = ϕ(k, kl)Blw( k) + (1 −ϕ(k, kl))Bsw( k),(1) where ϕ( k, kl) is a roll-off function. The parametric long-wave spectrum used in this study is from Elfouhaily et al.[2], while the short-wave spectrum is based on an equilibrium approach described by Kudryavtsev et al.[6] with the implementation as in Kudryavtsev et al. (2014)[8]. To take the relaxation and direction changes of long waves into account, we replace in a second run the Elfouhaily longwave spectra by those from a SWAN model run. We additionally apply a linear model to account for the effects of currents on the short-wave spectrum[9]. From here on we refer to both models simply as Elfouhaily and SWAN. 2.2 NRCS and Doppler The implementation of the RIM and DopRIM models from the monostatic case follow Kudryavtsev et al.[6] and Hansen et al.[3] The RIM accounts for three scattering mechanisms: specular scattering, Bragg scattering and scattering from wave breaking. This results in a monostatic NRCS as given by σm 0=σsp(1 −q) + σBr(1 −q) + σwbq, (2) where qthe fraction of the surface covered by breakers. The derivation of the bistatic NRCS is rather elaborate and is discussed in an upcoming publication (Kleinherenbrink et al. 2023, in preparation). A principal polarization basis as defined in [4] is used, which results in different weighting of the scattering mechanisms, which is summarized as σb 0=wb spσsp(1 −q) + wb BrσBr(1 −q) + wwbσb wbq, (3) where the weights depend on the bistatic geometry. We establish one description for the monostatic and bistatic Doppler fm,b D=σ′ spfsp +σHH′ Br fHH Br +σV V ′ Br fV V Br +σ′ wbfwb,(4) which depends on the scattering ratios σ′ xx where the superscript (b, m) is omitted for clarity. Due to the geometry, the bistatic Doppler can be described as a function of both HH and V V components (see Kleinherenbrink et al. (2023), in preparation) even though Sentinel-1’s transmission is in V V . For the monostatic case σHH′ Br becomes zero. 3 Results Fig. 1 shows the NRCS, in the major polarizations, computed from a run of a coupled ocean-atmosphere model near the coast of Calfornia. The top panels (based on the Elfouhaily long-wave spectrum) clearly show the directional dependence of backscatter. The line-of-sight or Harmony-A is more aligned with the wind than the line-of-sight of Harmony-B, resulting in a higher NRCS. Several linear and meandering features are visible that closely coincide with wind-speed anomalies and currents. (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure 1: Modelled observed NRCS (dB) in the major polarization near California using the output of a scientific workbench run. The top panels (a-c) are computed using an equilibrium approach with a Elfouhaily long-wave spectrum. The middle panels (d-f) are generated with SWAN as the long-wave spectrum. The bottom panels (g-i) show the normalized linear differences between (a-c) and (d-f) with in the background the sea surface temperature. The second row of panels shows the results of the SWAN model. The differences with the Elfouhaily model are shown in the bottom panels. The effects of currents on the wave spectrum are not explicitly modelled in the Elfouhaily model, which results in the largest discrepancies between the two models. The enhanced short-scale roughness near currents primarily result from enhanced wave breaking [8], which is primarily a local effect. Wind anomalies and currents can also introduce directional changes to the long-wave spectra, resulting in changes in the energy exchanges between long waves and short waves. Large-scale differences between the Elfouhaily and the SWAN model also arise from differences in wave age and (asymmetric) spreading leading to changes in short-wave energy balances. As differences in NRCS are typically in the order of 10-20% and vary with direction, there are consequence for the inversion of high-resolution stress-equivalent wind products. Fig. 2 shows the observed geophysical Doppler from the Elfouhaily and SWAN models. Whereas the NRCS is primarily regulated by short waves in the form of Bragg and wave breaking, the wave-Doppler results from line-of-sight-projected surface velocities of longer waves, O(10 m), and their tilt and hydrodynamic modulations of shorter waves [3]. The interpretation of the wave-Doppler is due to its dependence on both short and long waves even less trivial than the NRCS. A change in observed Doppler direction and magnitude might therefore arise from either a current, a change in the short-wave spectrum or a change in the long-wave spectrum. Longer waves are refracted by currents and slowly react to wind variations, as their relaxation scales are large O(10 km)[7]. Shorter waves react more rapidly to changing conditions and therefore change the tilt and hydrodynamic modulations. To help with the interpretation, or inversion, of Harmony’s observations at highresolution, a classification based on the scene’s features might therefore be required. 4 Conclusions At small scales, the observed NRCS and Doppler deviate substantially from stead-state conditions. Both multistatic radar observations are affected by currents, winds and depend on the inverse wave age. The retrieval of high-resolution stressequivalent wind and ocean-surface currents is therefore nontrivial and we cannot rely on traditional GMFs designed for low resolutions. Removal of the wave-Doppler based on solely stress-equivalent wind and inverse-wave age is not suitable for high resolutions, as other relaxation scales are involved. The inversion of both parameters requires an integrated approach involving multiple scales, SAR-spectral parameters and might even profit from an adjoint approach. References [1] N. Booij, R. C. Ris, and L. H. Holthuijsen. A thirdgeneration wave model for coastal regions: 1. Model description and validation. Journal of Geophysical Research: Oceans, 104(C4):7649–7666, Apr. 1999. [2] T. Elfouhaily, B. Chapron, K. Katsaros, and D. Vandemark. A unified directional spectrum for long and short winddriven waves. Journal of Geophysical Research: Oceans, 102(C7):15781–15796, July 1997. [3] M. Hansen, V. Kudryavtsev, B. Chapron, J. Johannessen, F. Collard, K.-F. Dagestad, and A. Mouche. Simulation of radar backscatter and Doppler shifts of wave–current interaction in the presence of strong tidal current. Remote Sensing of Environment, 120:113–122, May 2012. [4] L. Iannini, D. Comite, N. Pierdicca, and P. Lopez-Dekker. Rough-Surface Polarimetry in Companion SAR Missions. IEEE Transactions on Geoscience and Remote Sensing, 60:1–15, 2022. [5] J. A. Johannessen, B. Chapron, F. Collard, V. Kudryavtsev, A. Mouche, D. Akimov, and K.-F. Dagestad. Direct ocean surface velocity measurements from space: Improved quantitative interpretation of Envisat ASAR observations. Geophysical Research Letters, 35(22):L22608, Nov. 2008. [6] V. Kudryavtsev. A semiempirical model of the normalized radar cross section of the sea surface, 2. Radar modulation transfer function. Journal of Geophysical Research, 108(C3):8055, 2003. [7] V. Kudryavtsev. On radar imaging of current features: 1. Model and comparison with observations. Journal of Geophysical Research, 110(C7):C07016, 2005. [8] V. Kudryavtsev, B. Chapron, and V. Makin. Impact of wind waves on the air-sea fluxes: A coupled model. Journal of Geophysical Research: Oceans, 119(2):1217–1236, Feb. 2014. [9] N. Rascle, J. Molemaker, L. Mari´e, F. Nouguier, B. Chapron, B. Lund, and A. Mouche. Intense deformation field at oceanic front inferred from directional sea surface roughness observations: DIRECTIONAL ROUGHNESS AT OCEANIC FRONT. Geophysical Research Letters, 44(11):5599–5608, June 2017. (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure 2: Modelled observed Doppler (Hz) in the major polarization near California using the output of a scientific workbench run. The top panels (a-c) are computed using an equilibrium approach with a Elfouhaily long-wave spectrum. The middle panels (d-f) are generated with SWAN as the long-wave spectrum. The bottom panels (g-i) show the linear difference between (a-c) and (d-f) with in the background sea surface temperature. OBSERVATIONS OF THE AGULHAS CURRENT BY ALONG-TRACK INTERFEROMETRIC SYNTHETIC APERTURE RADAR Anis Elyouncha1, Leif E. B. Eriksson1, Christine Gommenginger2 1Chalmers University of Technology, Gothenburg, Sweden 2National Oceanography Center, Southampton, UK Synthetic aperture radar (SAR) offers the possibility to observe the sea surface circulation with very high spatial resolution. These observations are particularly relevant in coastal areas and shelf seas, where the ocean circulation is complex and highly variable. SAR has been routinely providing valuable information on sea surface winds and waves for many decades. During the last decade, a new application of SAR measurements based on the analysis of the Doppler shift has emerged [1, 2], opening the possibility to measure directly the surface currents. There are still however many unresolved questions and challenges. One of the challenging questions is the wind-wave-current interaction and its effect on the wind and wave retrieval. There are two major techniques for extracting ocean surface currents from SAR data: Along-track interferometry (ATI) and the Doppler centroid anomaly (DCA) analysis. The retrieval of surface velocity using these two techniques has been demonstrated in several papers, e.g. [3, 4, 1, 5]. A known effect that makes the retrieval of the ocean currents from SAR data challenging is the so-called wave-induced Doppler shift or velocity bias. This effect is due to the correlation between the normalized radar cross section (NRCS) and Doppler shift modulations, which are generated by the long modulating waves. Thomson [6] was probably the first to show that the Doppler spectrum was considerably shifted from the frequency of the Bragg waves. He attributed this to the fact that both the radar cross section and the radial surface velocity are functions of the position of the radar footprint along the wave. The Agulhas Current is the strongest western boundary currents in the southern hemisphere. The region of the Agulhas Current is characterized by a complex upper ocean dynamics involving a wide range of mesoscale and submesoscale processes. It thus provides an ideal natural laboratory for oceanographers and for testing remote sensing sensors and techniques. It is known that oceanic currents modify the wave filed properties, e.g. [7]. The Agulhas region has attracted the attention due to the rich interaction of the Agulhas Current system with wave field, e.g. [8]. This interaction is manifested in the SAR and optical images [8]. The manifestation of this interaction has has been reported by previous authors using the C-band SAR on-board Envisat [9] and Sentinel-1 [10] and also using Sentinel-2 imagery, e.g. [11]. In this paper the manifestation of the interaction of the Agulhas Current with the wind and the wave fields is presented and discussed based on the acquisitions provided by the X-band ATI-SAR TanDEM-X. To our knowledge this is the first time the Agulhas Current is mapped with an X-band spaceborne ATI-SAR. The backscatter and the Doppler shift derived from these unique acquisitions of the interferometric SAR over the Agulhas Current with very high spatial resolution (few hundreds of meters) are analyzed. Collocated existing products of ocean surface wind, ocean surface current, sea surface temperature and significant wave height are also analyzed to help the interpretation of these SAR observations. 1. PRELIMINARY RESULTS Figure 1 shows an example of the preliminary results from a TanDEM-X acquisition on 2015-11-05, together with model data for sea surface current and sea surface temperature (SST). Panels (a) and (b) depict the observed NRCS and Doppler centroid (DC), respectively. These observations contain complex signatures of the varying wind and current fields in addition to surface and internal waves. The DC image captures accurately the boundary and the intensity of the Agulhas Current. Moreover, these maps show unprecedented fine structure of the Agulhas Current and its interaction with the wave field. The core of the Agulhas Current is clearly detected by a increase in the DC image around -35 degrees latitude. The pattern depicted by the backscatter images is on the other hand much more dependent on the wind and sea state than on the current velocity. This is manifested by a sharp enhancement of the NRCS at the northern edge of the current. The structure of the current core can also be discerned by increase in the NRCS image around the same latitude as of the DC maximum. Panel (c) depicts the model current speed. It can be observed that the northern boundary around -34.5 degrees latitude depicted by the NRCS and DC images is unresolved by the model. This is probably an indication that it is due to a wind front rather than an oceanic front. The wind field will be analyzed later to investigate this. Panel (d) depicts the azimuth profile of the model current speed, SAR-derived radial velocity and the sea surface temperature. The model current and SST are provided by the Copernicus product GLOBAL MULTIYEAR PHY 001 030. It can be observed that the SAR-derived velocity locates correctly the current core which reaches 1.5 m/s. The model current and SST both peak roughly at the same location but the variation is much more smoother than SAR observation. The influence of the Agulhas Current on the wind and wave field retrieval will be investigated. The inversion of the backscatter to wind speed, without taking the current into account, would lead artificially high estimates of SAR-derived wind speeds in the regions of enhanced backscatter. Note that the effect of the current on the wind and wave field retrieval will also impact the current retrieval via the wave-induced Doppler shift. To summarize, this paper analyses collocated NRCS and DC images acquired by the along-track interferometric SAR TanDEM-X over the Agulhas Current. These observation present rich manifestations of the wave-current and wind-current interaction at very high resolutions not observed before from satellites. The roughness and velocity signatures depend on the wind, wave and the current fields. A case of particularly enhanced roughness, which is probably due to a convergent wind front is reported. More detailed investigation of the cause of this sharp enhancement of the NRCS in regions of strong current shear will be carried out. Finally, more cases from other TanDEM-X acquisitions will be shown at the conference. 2. REFERENCES [1] B. Chapron, F. Collard, and F. Ardhuin, “Direct measurements of ocean surface velocity from space: Interpretation and validation,” Journal of Geophysical Research, vol. 110, Mar 2005. [2] R. Romeiser and D. R. Thompson, “Numerical study on the along-track interferometric radar imaging mechanism of oceanic surface currents,” IEEE Transactions on Geoscience and Remote Sensing. Vol. 38-II, 446-458, 2000., vol. 38, no. 2, pp. 446–458, Jan 2000. [3] R. Romeiser, H. Runge, S. Suchandt, R. Kahle, C. Rossi, and P. S. Bell, “Quality assessment of surface current fields from TerraSAR-X and TanDEM-X along-track interferometry and doppler centroid analysis,” IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 5, pp. 2759–2772, 2014. [4] A. Elyouncha, L. E. B. Eriksson, R. Romeiser, and L. M. H. Ulander, “Measurements of sea surface currents in the Baltic Sea region using spaceborne along-track InSAR,” IEEE Transactions on Geoscience and Remote Sensing, vol. 57, no. 11, pp. 8584–8599, Nov 2019. [5] J. A. Johannessen, B. Chapron, F. Collard, V. Kudryavtsev, A. Mouche, D. Akimov, and K.F. Dagestad, “Direct ocean surface velocity measurements from space: Improved quantitative interpretation of Envisat ASAR observations,” Geophysical Research Letters, vol. 35, no. 22, 2008. [6] D. R. Thompson and J. R. Jensen, “Synthetic aperture radar interferometry applied to ship generated internal waves in the 1989 Loch Linnhe experiment,” Journal of Geophysical Research: Oceans, vol. 98, no. C6, pp. 10259–10269, 1993. [7] Fabrice Ardhuin, Sarah T. Gille, Dimitris Menemenlis, Cesar B. Rocha, Nicolas Rascle, Bertrand Chapron, Jonathan Gula, and Jeroen Molemaker, “Small-scale open ocean currents have large effects on wind wave heights,” Journal of Geophysical Research: Oceans, vol. 122, no. 6, pp. 4500–4517, 2017. [8] Vladimir Kudryavtsev, Alexander Myasoedov, Bertrand Chapron, Johnny A. Johannessen, and Fabrice Collard, “Imaging mesoscale upper ocean dynamics using synthetic aperture radar and optical data,” Journal of Geophysical Research: Oceans, vol. 117, no. C4, 2012. [9] M. Krug, A. Mouche, F. Collard, J. A. Johannessen, and B. Chapron, “Mapping the agulhas current from space: An assessment of asar surface current velocities,” Journal of Geophysical Research: Oceans, vol. 115, no. C10, 2010. [10] Y. Quilfen, M. Yurovskaya, B. Chapron, and F. Ardhuin, “Storm waves focusing and steepening in the agulhas current: Satellite observations and modeling,” Remote Sensing of Environment, vol. 216, pp. 561–571, 2018. [11] Vladimir Kudryavtsev, Maria Yurovskaya, Bertrand Chapron, Fabrice Collard, and Craig Donlon, “Sun glitter imagery of surface waves. part 2: Waves transformation on ocean currents,” Journal of Geophysical Research: Oceans, vol. 122, no. 2, pp. 1384–1399, 2017. Figure 3: (background) NovaSAR roughness image in S-band at the time of OSCAR acquisition. (in color) OSCAR retrieved 2D current along the track. Near range is on the East part of the image. ACKNOWLEDGEMENTS This work was supported by ESA/ESTEC Contract Number 4000116410/16/NL/BJ for the OSCAR development and ESA/ESTEC contract number 400017623/22/NL/IA for the campaign over Iroise Sea. REFERENCES [1] Adrien C.H. Martin, Christine Gommenginger , “Towards wide-swath high-resolution mapping of total ocean surface current vectors from space: Airborne proof-of-concept and validation”, Remote Sensing of Environment, Volume 197, 2017, Pages 58-71 [2] Adrien C.H. Martin, Christine Gommenginger, Jose Marquez, Sam Doody, Victor Navarro, Christopher Buck, “Wind-wave-induced velocity in ATI SAR ocean surface currents: First experimental evidence from an airborne campaign”, Journal of Geophysical Research: Oceans, Volume121, Issue3, March 2016, Pages 1640-1653 [3] Gommenginger, Christine et al. , “SEASTAR: A Mission to Study Ocean Submesoscale Dynamics and Small-Scale AtmosphereOcean Processes in Coastal, Shelf and Polar Seas,” Frontiers in Marine Science, Volume 6, p.457, 2019. Ocean Surface Currents from Sentinel-1 Doppler observations Artem Moiseev1, Fabrice Collard2, Johnny A. Johannessen1,3 1Nansen Environmental and Remote Sensing Center (NERSC), Bergen, Norway 2OceanDataLab (ODL), Locmaria-Plouzané, France 3Geophysical Institute, University of Bergen, Norway Observations from satellite Synthetic Aperture Radars (SARs) can complement existing ocean observing systems with systematic kilometer-resolution observations of the total surface radial current velocity in the open ocean and coastal zones. Reliable and systematic observations of the total ocean surface currents are essential for monitoring, model validation, search and rescue operations, pollution applications, and climate research. However, such observations are challenging and expensive to acquire and, hence, rarely available on a systematic basis. In this study focusing on the greater Agulhas Current, the surface current retrieval algorithm for Sentinel-1 developed and validated in the ESA WOC project is presented. We retrieved surface current radial velocities from the SAR Doppler Centroid Anomaly (DCA). This DCA is derived from the frequency difference between transmitted and backscattered signals after correcting for the satellite motion (geometric Doppler). Over the ocean, the DCA is a measure of the surface motion induced by the combination of the wind and waves (sea state) and ocean surface currents in the antenna line-of-sight direction. The sea-state-induced contribution to the DCA can be estimated from Empirical Geophysical Model Functions (GMFs), such as CDOP3SiX, for given radar incidence and polarization, based on wind and wave fields from collocated models. We evaluated the impact on the DCA of different global wind (ECMWF, NCEP GFS, ERA5) and wave (MFWAM, WAVERYS, WW3) model forecasts and reanalysis. Although usage of the model fields is convenient due to their regular availability, it also yields limitations related to the accuracy and representativeness of the model product which will strongly impact the estimate of the sea state contribution to the DCA and, consequently, the SAR-derived radial surface current retrievals. As such we explore the potential of using SAR-derived wind and wave information to estimate the sea state DCA. We used wind fields routinely provided in Sentinel-1 L2 OWI product. For the wave field, we tested Sentinel-1 L2 OSW product as well as cross-spectra estimates provided for the Wave (WV) mode acquisitions. The new product was then compared with the previous product (see Figure 1b). Moreover, it was compared with the QG current product derived from altimetry (Figure 1c). Both SARand altimetry-derived surface current products agree on the location of the main ocean surface current features. We also demonstrated the potential of using the Sentinel-1-derived ocean surface current to validate model forecasts in the region. We collocated the SAR RVLs with ocean model fields from Mercator 1/12 deg. model (available from CMEMS) along the southern African coast. The model can reproduce the Agulhas Current features, however, disagreement with observations in terms of the location and velocity of meanders and eddies are evident. In summary, we have demonstrated that the Sentinel-1 acquisition can provide valuable observations of the ocean surface current which are essential for systematic model validation, applications, and climate research. We also demonstrated the potential of using SAR-derived surface currents for model validation. The accuracy of the ocean surface current radial velocity retrieved from SAR, on the other hand, relies on the precise removal of the sea-state-induced signal. We developed and demonstrated a new approach for estimating this sea state contribution based on the wind and wave information which can be directly extracted from SAR observations. This approach provides a more reliable way to estimate sea state DCA compared to using auxiliary model fields. However, the algorithm requires more testing and systematic validation. The developed methodology is promising and ready to be applied to the operational Sentinel-1 mission (sustained operation until 2030) as well as for candidate future satellite missions designed for monitoring of the upper ocean circulation (e.g., SEASTAR and Harmony). Figure 1. The Sentinel-1A SAR scene acquired in ascending pass on 7 July 2019 at 16:36:36: (a) Total surface radial velocity (i.e., waveand surface-current-induced signal) from the Doppler shift; (b) Ocean surface current radial velocity (i.e., after removal of the sea-state-induced signal) from the Doppler shift; (c) Normalized radar cross-section. The blue/red color in a, b indicates south-westward/north-eastward velocity. The SAR scene is collocated with: (d) Wind speed (color) and direction (arrows) at 10m height from the ECMFW; (e) Geostrophic velocity from altimetry observations; (f) Satellite-derived Sea Surface Temperature field. The black frames indicate the footprint of the SAR acquisition frames. The dashed contours in subplots b, c, e, f represent the position of the cyclonic (CE) and anticyclonic (ACE) eddies detected in the geostrophic velocity field Towards Calibrated Sentinel-1 OCN RVL Products G. Guitton1, F. Collard1, H. Johnsen2, G. Engen2, A. Recchia3, A. Cotrufo3, S. Brás4, N. Miranda5, M. Pinheiro5 1OceanDataLab, Locmaria, Plouzane, France. 2NORCE Norwegian Research Center AS, Tromsø, Norway. 3Aresys Ltd, Milano, Italy 4ESTEC/ESA, Noordwijk, The Netherlands 5ESRIN/ESA, Frascati, Italy Corresponding author: G. Guitton ([email protected]) Co-author: F. Collard (dr . fab @oceandatalab.com ), H. Johnsen ([email protected]), G. Engen ([email protected]), A. Recchia ([email protected]), A. Cotrufo ([email protected]), M. Pinheiro ([email protected]), N. Miranda ([email protected]), S. Brás ([email protected]) Abstract The Doppler Centroid (DC) frequency shift recorded over ocean surfaces by Synthetic Aperture Radar (SAR) is a sum of contributions from satellite attitude/antenna and ocean surface motion induced by waves and underlying ocean currents. A precise calibration of the DC is needed in order to predict and subsequently remove contributions from attitude/antenna. Recently, a novel data calibration technique based on combining gyroscope telemetry data and global Sentinel-1 WV OCN products has demonstrated promising capabilities to quantify the Sentinel-1 (S1) attitude and hence provide calibrated estimates of the corresponding DC frequency shift. For validation purpose, one year of S1 a and b WV OCN products, orbit data and gyroscope data are combined providing one year of restituted attitude data (AUX_ESTATT). For the same time period, mean DC bias versus elevation angle is computed on a daily basis from S1 IW land acquisitions (AUX_DCBIAS). The AUX_ESTATT and AUX_DCBIAS products are subsequently used to generate global data set of calibrated S1 WV OCN products as well as subsets of calibrated S1 IW OCN products from predefined super sites (Norwegian Coast, Agulhas, Mediterranean). In this paper we assess the accuracy and precision of the calibrated DC frequency of S1 WV and IW acquisitions acquired over both land and ocean areas. The DC standard deviation (STD) and bias show significant reduction for both satellites and for all swaths. Assessment of the performance of global WV data shows a STD around or less than 6Hz, while the BIAS is less than 2 Hz. The performance is very similar for both satellites and for both swaths. For IW the STD is similar, but small DC bias between sub-swaths is sometimes observed. The remaining errors are mainly due to change in antenna characteristics on a timescale not captured with the procedure used to generate the mean DC bias stored in the AUX_DCBIAS file. Such changes may come from thermo-elastic effects and/or temperature compensations applied to the antenna. This directly affects the IW mode DC, where it is also clearly visible in some scenes. For WV mode it mainly impacts the statistics. S1 WV mode has achieved a performance (i.e. accuracy and precision) within the requirement for climatology mapping of global ocean current features. It will be demonstrated on a longer time serie than the validation year. For IW mode, we have achieved a precision within the requirement, but use of land areas within the scene is still required to achieve the required accuracy over all sub-swaths. Improved accuracy using DC over nearest land will be demonstrated and the limitations discussed. Remaining issues related to antenna temperature compensation and Level 0 Doppler estimates will also be discussed. Figure: Histograms of observed (---) and calibrated (___) DC from S1a WV (left) and S1a IW (right) modes. Data acquired over land areas within the period 01 May 2019 to 30 April 2020. EVALUATION OF SURFACE CURRENTS DERIVED FROM SENTINEL-1 SAR DOPPLER SHIFT IN THE NORTHWESTERN MEDITERRANEAN SEA USING COASTAL HF RADARS Baptiste Domps Degreane Horizon Cuers, France [email protected] Charles-Antoine Gu´ erin Univ. Toulon Aix-Marseille Univ., CNRS, IRD, MIO Toulon, France [email protected] ABSTRACT We evaluate the estimations of oceanic surface currents retrieved from C-band synthetic aperture radar onboard Sentinel-1A/B. The assessment is carried out using reference measurements provided by a network of coastal highfrequency radars on the French Rivieira over 6 months. 1. INTRODUCTION Real-time observations of coastal ocean surface currents have been routinely performed by high-frequency radars (HFR) for nearly half a century, starting with the pioneering works of Crombie [1] and Barrick [2]. Nowadays, the vast majority of HFR operate from shore in the HF and UHF frequency bands, typically from 5 to 50MHz. Within this range, commercial HFR are proved to provide reliable estimates of surface currents at high spatial (typically 5km) and temporal (typically 30min) resolution, up to 100km from coast [3]. The Mediterranean Institute of Oceanography (MIO) and the University of Toulon have been operating for more than one decade a network of such systems, manufactured by Helzel Messtechnik GmbH [4]. The network is located on the Western French Riviera, in the vicinity of Toulon, and consists in an original arrangement of spatially diverse systems with 2 distinct transmitters and receivers on 3 remote sites [5]. Such multistatic geometry allows to derive the total vector surface current on a1 km×1 km Cartesian grid cell every 20min whose quality have been assessed through two campaigns of in situ measurements [6]. Besides a wide range of oceanographic and operationnal applications, the Toulon HFR network is also an experimental ground for developing new signal processing techniques [7, 8]. Aside from ground-based systems, the measurement of the total suface current velocity from space at high resolution remains a challenge partially met by synthetic aperture radars (SAR) such as ASAR (formerly onboard ENVISAT) and CSAR (currently onboard Sentinel-1). It is well known that SAR does not only measure the magnitude of the backscattered signal but also the Doppler frequency offset induced by the velocity of the scatterers. However, the Doppler shift is not simply proportionnal to the line-of-sight surface current. The surface current actually represents a minor contribution to the total frequency anomaly. The measurement is affected by a number of instrumental and environmental biases. While the former can be withdrawn by calibrating the frequency measurements (see, e.g., [9, 10]), doing so is less straightforward for the latter. The geophysical Doppler shift is yet dominated by wave-induced surface displacements whose contribution to the frequency anomaly can not be evaluated directly from the SAR measurements. One popular way to determine the wave-induced Doppler bias is to use a geophysical modeling function (GMF) such as CDOP [11]. This empirical GMF relates the wave bias to the surface wind speed and direction than can be routinely obtained from SAR measurements. As of now the use of surface currents inferred from SAR measurements has not received a place within the oceanographic community. Such current estimations yet suffer from a lack of evaluation and uncertainty analysis. In the past few years, there have been isolated atempts to assess the validity of surface currents retrieved from Sentinel-1A/B using coastal HF radars [12, 13]. The results of these studies are promising and we contribute to this effort by using reference estimates provided by the HF radar network of Toulon. 2. EVALUATION OF SURFACE CURRENTS Sentinel-1A/B measurements have been processed using the standard scheme (we refer to Fig. 1 of [9] for a description). The vector surface currents provided by the HFR network have been interpolated along the Sentinel-1 cartesian grid and projected along the SAR line-of-sight. An example of comparison between S1A/B and HFR currents is given on Figs. 1c and f. After visual inspection of the whole dataset, a total of 6 months of colocalized measurements have been compared from July 2020 to March 2021, accounting for about 150,000 individual comparisons. Extended abstract submitted to SEASAR 2023 42.3°N 42.45°N 42.6°N 42.75°N 42.9°N 43.05°N 0 25 50 75 (a) cm·s−1 U0 5 10 (b) m·s−1 W−75 0 75 (c) cm·s−1 Ur,HFR 5.6°E 5.8°E 6°E 6.2°E 6.4°E 6.6°E 42.3°N 42.45°N 42.6°N 42.75°N 42.9°N 43.05°N −20 −15 −10 (d) dBσ0 5.6°E 5.8°E 6°E 6.2°E 6.4°E 6.6°E 16 18 20 22 24 (e) cm·s−1 CDOP 5.6°E 5.8°E 6°E 6.2°E 6.4°E 6.6°E −75 0 75 (f) Hz Ur,S1A Fig. 1. Maps (colorscale) obtained off of Toulon on March 1st, 2021, at 17:30 UTC: (a) 2D surface current measured by the coastal HFR network of Toulon; (b) 2D 10-m wind obtained using CMOD7; (c) HFR surface current projected along S1A line-of-sight; (d) SAR NRCS; (e) wind-wave Doppler bias computed with CDOP; (f) radial surface current obtained from S1A measurements [14]. 2.1. Overall Results Fig. 2 shows the estimated surface velocity inferred from S1A/B SAR versus the projected surface velocity retrieved by the HFR network of Toulon. The root mean square error (RMSE) is of 21cms−1. Thorough comparisons between ascending and descending orbits, Sentinel-1A and -1B or surface wind velocities will be presented during the conference. −0 5 0 0 5 −0 5 0 100 101 102 Projected HFR velocity (m.s-1) 0.5 . . S1A/B velocity (m.s-1) No. of estimates . ---- y=x —– y= 0,71x+ 0,02 RMSE = 21 cm s−1 R2= 0,53 Fig. 2. Scatter plot (color scale, counts) of estimated surface velocities using the HFR network versus S1A/B. 2.2. Wave-induced bias Within the classical processing scheme for inverting the radial surface current from the SAR Doppler shift, the waveinduced Doppler bias is given by a model (such as CDOP) as a function of the relative wind speed and direction fww =CDOPVV(u10, θ10)(1) and is then substracted from the measured Doppler anomaly to give the current-induced Doppler shift fc=fdca −fww (2) where fc= 2Ur/λ0. Here we use the surface current velocity provided by the HFR network as an a priori information for the frequency shift, plus the wind information retrieved from the SAR NRCS, to retrieve the wave-induced bias. This observed bias is then used to train a simple neural network, CDOP-TLN, as an alternative to CDOP suited for the observations in the vicinity of Toulon. This simple, empirical model yields to better estimates of the wave-induced bias than CDOP (Fig. 3), being however only valid in the coastal Mediterranean sea. 3. CONCLUSIONS AND FUTURE WORK This study shows a remarkable agreement between the estimations of surface currents provided by Sentinel-1A/B and the HFR network of Toulon. Our ongoing research focuses on the thorough evaluation of these estimations. Furthermore, 2 Copernicus Marine Service SITAC SAR-Based Baltic Sea Ice Products Finnish Meteorological Institute (FMI) is providing automated sea ice products as part of the Copernicus Marine Service (CMS, part of the European Commission Copernicus Programme) sea ice thematic assembly center (SITAC) coordinated by MET Norway. The spatial cover of the products is the whole Baltic Sea and a typical temporal coverage is 1-3 days, depending on the availablility of SAR imagery. In the presentation an overview of the automated FMI SAR-based Baltic Sea ice products, their processing chain and their estimated quality will be presented, including examples of the products. Also plans for improvements and future new products will be discussed. The current products estimate three essential sea ice parameters: sea ice thickness (SIT), sea ice concentration (SIC) and sea ice drift (SID). During each Baltic Sea ice season, lasting approximately from the beginning of December until late May hundreds of SAR-based sea ice products are delivered to CMS by FMI. In Figure 1 the monthly amounts of delivered products grouped by product category for the season 2021-2023 are shown. Figure 1. FMI CMS SITAC Baltic Sea ice products during the season 2021-2022 by category. The IC SIC and IC SIT are based on the daily ice charts made by the ice analysts at FMI/SMHI, the other products are automated and based mainly on SAR data. The data quality is evaluated after each season and the bias, L1 difference (L1D) and root-meansquare difference (RMSD) for the estimates with respect to reference data sets are provided. Sea ice Thickness (SIT) SIT is currently based on SAR imagery supported by background information of the previous day ice chart. From SAR imagery only qualitative SIT information can be derived. For this reason quantitative background information of SIT is required. The current operational algorithm (Karvonen et al., 2003) uses ice chart ice thickness as its background information. It has also been shown that the background information can be produced by an ice model (Karvonen et al. 2008). The operational SIT algorithm was developed for C-band HH polarized data and it is applied to Sentinel-1 extra wide swath (EW) GRDM mode HH/HV-polarized data and to Radarsat-2 wide swath mode (WSM) HH/HV data. A slightly modified SIT algorithm is applied to Sentinel-1 IW mode VV/VH polarized data. The SIT products are given in 500m resolution and provided to the CMS service in near-real-time (NRT), CMS time requirement being four hours from the SAR acquisition. In addition to the SIT grids, corresponding to each SAR image, a daily SIT mosaic is provided. In the daily SIT mosaic the most recent SIT information is overlaid over the older information and each grid point then contains the most recent SIT information available at the moment of the mosaic generation. The reference data in evaluation of the SIT estimates are SIT measurements made by the Finnish and Swedish ice breakers. In the 2021-2022 evaluation the bias for the EW/WSM data was -0.7 cm (slight underestimation), L1D 8.5 cm and RMSD 10.2 cm. For the IW mode data the corresponding measures were 3.2 cm, 11.2 cm and 15.3 cm, indicating that SIT estimation using VV/VH data has slightly worse estimation accuracy than HH/HV data. The SIT estimation accuracy during 20212022 was similar as in the previous seasonal evaluations. Sea Ice Concentration (SIC) SIC is the fraction of sea ice within an area, which can be a grid point area or a segment defined by SAR segmentation or polygon drawn by hand. In the FMI SAR based products the reference areas are SAR segments produced by a segmentation algorithm. The algorithm provides SIC in percents, i.e. 0 % represents all open water and 100% all sea ice over the reference area. Baltic Sea SIC is provided daily in the morning after the Radarsat-2 and Sentinel-1 morning passes over the Baltic Sea. The product is a daily mosaic and generated from single image SIC grids in the same manner as the SIT mosaic. SIC is provided in 500m resolution. The SIC algorithm currently uses C-band HH(HV-polarized SAR data from Sentinel-1 and Radarsat-2 an additionally also microwave radiometer data from AMSR2. The algorithm is based on a multilayer perceptron (MLP) neural network combining texture measure inputs based on the SAR data and polarization and gradient ratios based on the AMSR2 brightness temperatures of different frequency channels. The algorithm is described in detail in (Karvonen, 2017). In evaluation of the SIC data, the FMI ice chart SIC and 3.125 km resolution ASI algorithm (Spreen et al. 2008) SIC, also based on AMSR2 data, are used as reference SIT data sets. For the season 2021-2022 data the bias, L1D and RMSD were -3.7 (underestimation), 10.4 and 26.2 percentage points, respectively. The corresponding values when compared to ice chart SIC were -1.0, 7.0 and 22.3 percentage points. Also the 2021-2022 SIC estimation accuracy has remained similar as during the previous seasons. Sea Ice Drift (SID) SID is estimated based on matching two SAR images acquired at different time instants over the same area, i.e. based on multitemporal SAR analysis. Sentinel-1 EW mode and Radarsat-2 SCW data are currently used in the SID estimation. Because during longer time periods sea ice may deform too much for reliable matching of the multitemporal SAR data, we restrict the time difference between the SAR images to be less than three days. Baltic Sea SID was for a long time estimated using a two-resolution phase correlation approach (Karvonen 2012). However, since 2020 a new algorithm was established. The new algorithm also operates in two resolutions, first a coarse resolution pattern matching by the ORB algorithm (Rublee et al. 2011) is performed and the resulting coarse resolution SID is then refined in fine resolution by applying Lucas-Kanade optical flow (Lucas and Kanade 1981). Optical flow has the capability of estimating the movement from an image to another with a sub-pixel resolution. SID estimates are evaluated using ice drifter buoys deployed almost every winter. The evaluation is performed separately for short drift and larger drift. For larger drift both direction and magnitude are evaluated, for short drift only magnitude. In general, the estimated drift and buoy drift correspond to each other quite well. However, there exist some outliers that typically give short SID estimates for some significantly longer buoy motions. It seems that these cases occur near the boundary of static and drift ice. Product development We are continually developing the products and try to involve new SAR data in the production chain. Currently the integration or Radarsat Constellation Mission (RCM) HH/HV C-band SAR data is under construction. The RCM rdata reireval and preprocessing have already been implemented but RCM data are not used for the products yet. Also X-band SIT algorithm utilizing X-band HH-polarized SAR data from TerraSAR-X, Cosmo-SkyMED and PAZ is currently in operational test phase. We are also studying good ways to estimate the degree of ice deformation and identifying single large pressure ridges based on SAR imagery. We expect that this work will later result into a SAR degree of deformation classification product to be included in the FMI CMS SITAC product portfolio. To improve estimation accuracy we have plans to utilize convolutional neural netrorks (CNN) in SIC estimation and possibly for other sea ice parameters as well. We have already successfully applied CNN to Sentinel-1 SAR data for SIC estimation (Karvonen 2022) and integrating AMSR2 microwave radiometer data in the CNN model is well under construction. References J. Karvonen, M. Simila, I. Heiler, Ice Thickness Estimation Using SAR Data and Ice Thickness History, Proceedings of the IEEE International Geoscience and Remote Sensing Symposium 2003 (IGARSS'03), 2003. J. Karvonen B. Cheng, M. Simila, Ice Thickness Charts Produced by C-Band SAR Imagery and HIGHTSI Thermodynamic Ice Model, Proc. of the Sixth Workshop on Baltic Sea Ice Climate, pp. 71–81, Lammi, Finland , 2008. J. Karvonen, Operational SAR-based sea ice drift monitoring over the Baltic Sea, Ocean Sci., 8, 473-483, doi:10.5194/os-8-473-2012, 2012. http://www.ocean-sci.net/8/473/2012/os-8-473-2012.html J. Karvonen, Baltic Sea Ice Concentration Estimation Using SENTINEL-1 SAR and AMSR2 Microwave Radiometer Data, IEEE Transactions on Geoscience and Remote Sensing, v. 55, n. 5, pp. 2871 - 2883, 2017. J. Karvonen, Baltic Sea Ice Concentration Estimation From C-Band Dual-Polarized SAR Imagery by Image Segmentation and Convolutional Neural Networks, in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-11, 2022, Art no. 4301411, doi: 10.1109/TGRS.2021.3097885, 2022. B. D. Lucas and T. Kanade, An iterative image registration technique with an application to stereo vision. Proceedings of Imaging Understanding Workshop, pages 121–130, 1981. E. Rublee, V. Rabaud, K. Konolige and G. Bradski, ORB: An efficient alternative to SIFT or SURF, 2011 International Conference on Computer Vision, Barcelona, Spain, pp. 2564-2571, doi: 10.1109/ICCV.2011.6126544, 2011. Spreen, G., L. Kaleschke, and G.Heygster (2008), Sea ice remote sensing using AMSR-E 89 GHz channels J. Geophys. Res.,vol. 113, C02S03, doi:10.1029/2005JC003384. 1 Combining CAnd L-band SAR Imagery For Automated Sea Ice Classification And Segmentation Johannes Lohse1and Wolfgang Dierking1,2 1) UiT The Arctic University of Norway 2) Alfred Wegener Institute, Helmholtz Center for Polar and Marine Research Abstract Operational activities and environmental studies in the Arctic both require robust and reliable mapping of sea ice conditions. Because of its 24/7 imaging capability, synthetic aperture radar (SAR) is the main data source for operational ice charting in national ice services worldwide. At present, the daily distributed ice charts are largely produced manually by visual interpretation of the SAR imagery. The most commonly used frequency is C-band, and data is available for example through the Sentinel-1 or Radarsat-2 satellites. Future missions, such as for example ROSE-L or Sentinel-1 NG, are expected to further increase the total amount of available SAR images and will in addition provide data at different frequencies, in particular L-band, for operational monitoring. The huge amount of data will require (semi-)automated analysis, and studies are needed in order to determine the ideal configuration of future satellites missions. Here, we use a set of 161 aligned image pairs acquired at Cand L-band to investigate the benefits of combining both frequencies for multi-frequency classification and segmentation of sea ice types. The image pairs were acquired over three different test sites, located in Fram Strait, Lincoln Sea, and Belgica Bank. Sea ice drift between image acquisitions is compensated, using an algorithm developed at Chalmers University of Technology, such that we can stack the Cand L-band data pixel-by-pixel. We then perform supervised classification and unsupervised segmentation of the image pairs, using both single-frequency (Cand L-band as stand-alone) and dual-frequency approaches. We evaluate the classification results in terms of classification accuracy, and the segmentation results in terms of how many statistically separable clusters are found in each data set. Our results presented here clearly show the benefits of combining the complementary information from both frequencies, in particular for the classification of young ice, open water, and newly formed ice in lead areas. I. Introduction Monitoring and continuous mapping of the Arctic sea ice is essential to support marine traffic and navigation in the Arctic and to assess the state of the sea ice cover for environmental and climate studies. Hence, daily sea ice charts are provided by national ice services around the world. Because of its independence of sunlight and cloud conditions, synthetic aperture radar (SAR) is the main data source for operational ice charting. Production of the ice charts is at present done manually via visual interpretation of the imagery by expert ice analysts; a time-consuming process that depends on the training and experience of the individual analyst. Automation or semi-automation of the ice chart production (”computer-assisted ice charting”) is thus a desirable goal. Furthermore, the operational workflow in many ice services is predominantly based on imagery acquired at C-band, even though it has been shown that other frequencies, such as for example L-band, can offer valuable complementary information [1]. This focus on C-band is mostly caused by the fact that L-band data is not yet routinely available for operational purposes. However, by the end of this decade, ESA’s upcoming ROSE-L mission [2] will consistently provide L-band data which can be exploited for ice charting. The work presented here is part of the ESA project Synergistic Use of Land C-Band SAR Satellites for Sea Ice Monitoring (LC-ICE). In this study, we investigated the benefits of combining Cand L-band SAR imagery in a dualfrequency approach for automated mapping of sea ice types. This type of analysis requires perfectly aligned SAR data from both frequencies; the data set that we used is introduced in Section II. In Section III we briefly describe the algorithms and the layout of our experiment. We summarize and discuss the main results in Section IV. II. Data Set Multi-frequency classification and segmentation of sea ice imagery requires spatially overlapping image pairs at Cand L-band, acquired close in time. For the LC-ICE project, such image pairs were collected over different test sites in the Arctic. Here, we use data from the Belgica Bank,Fram Strait, and Lincoln Sea. The C-band images are acquired by Sentinel-1 (S1) and the L-band data by ALOS-2 PALSAR, respectively. All S1 data used in this study are in extra-wide swath (EW) mode. We use the ground range detected format at medium resolution (GRDM), which at present is the most commonly used S1 data format in operational sea ice charting. Our ALOS-2 data is available in both wide beam (WB) and fine beam (FB) mode. Specifications of the sensor acquisition modes used in this study are listed in Table 1. 2 TABLE I Sensor acquisition modes used in this study S1 EW GRDM ALOS-2 WB ALOS-2 FB swath width (km) 410 350 70 pixel spacing (m) 40x40 25x25 6.25x6.25 looks ENL: 10.7 5(r)x3(a) 2 polarizations HH, HV HH, HV HH, HV For a joint, multi-frequency analysis, the individual images in each image pair must be perfectly aligned, so that they can be stacked to a multi-dimensional data cube. Here, we use a set of 161 aligned image pairs, for which the sea ice drift during the time interval between the Cand L-band acquisition was compensated using an algorithm developed at Chalmers University of Technology [3]. The aligned image pairs are calibrated, noise-corrected, and geocoded; the backscatter intensities are converted to decibel (dB). III. Method A. Selection of areas of interest within image pairs We have visually inspected all image pairs to ensure correct alignment and to check for undesired alignment artifacts in the imagery. We have then selected an area of interest (AOI) within each pair that is suitable for the study of multi-frequency classification. The main goal of the AOI selection is to choose areas that are as large as possible, but small enough to: •avoid too wide IA range •avoid visible swath boundaries and influence of sensor noise •exclude areas with mis-alignment or alignment artifacts B. Classification and segmentation We apply both supervised classification and unsupervised segmentation on the selected AOIs, using algorithms developed by Lohse et al. (2020) [4] and Doulgeris (2015) [5], respectively. Both algorithms are based on well-established statistical methods that use multi-variate probability density functions and Bayesian decision rules. The classification algorithm assigns each image pixel to a set of class labels that are defined based on training data. We evaluate its performance in terms of classification accuracy (CA), which is estimated from a set of validation data. Regions of interest (ROIs) of different ice types for training and validation are selected manually based on visual interpretation of the combined C-/L-band image pairs. The segmentation algorithm automatically determines the number of clusters based on the statistics of the input data. As we gradually increase the sensitivity of the algorithm, the number of resulting clusters increases. We then use the Jeffries-Matusita (JM) distance to asses the separability of the clusters. Finally thresholding the JM distance allows us to find the optimal number of clusters and thus quantify the statistically separable information in the imagery in an unsupervised approach. We perform both classification and segmentation using single-frequency (C-band and L-band as stand-alone) and dual-frequency data as input, which allows us to assess the added value of combining both frequencies. IV. Results and Discussion Our results clearly show the benefit of combining Cand L-band data for automated separation of sea ice types. With the segmentation approach, the combination of Cand L-band on average separates 2.4 more clusters than Cband alone, and 1.0 more clusters than L-band alone. For the classification, we find that the dual-frequency approach achieves the highest CA in 98.1 % of all cases. In some cases, in particular for regions of deformed ice, the L-band stand-alone classification performs almost or equally as well as the dual-frequency classification. However, especially in lead areas with open water, newly formed ice, and young ice, the combination of Cand L-band is usually superior to either of the single-frequency approaches. An example AOI with selected ice type ROIs and a comparison of classification results is shown in Figure 1. It demonstrates the superiority of L-band for separating level and deformed ice compared to C-band; the good separation of these ice types at L-band is retained in the multi-frequency scenario. Furthermore, a lead stretching from east to west in the lower part of the image is generally identified in the stand-alone classification results at both C-band and L-band. However, at L-band, almost the entire lead is classified as Young Ice, while at C-band the lead is split up into OW/New Ice,Young Ice, and Level Ice. Only the combination of both frequencies manages to capture the lead as a mixture of OW/New Ice and Young Ice, with small parts in the northern area being open, and the majority of the lead being re-frozen. At the conference, we will present further examples with corresponding CAs and discuss differences between Cand L-band. 3 Fig. 1. Classification example from Fram Strait test site, 2019/12/10, acquisition time difference ∆t= 1.7hours. Cand L-band false-color RGB images (R-HV, G-HH, B-HH) with training and validation ROIs for different ice types are shown at the top, singleand multi-frequency classification results below, including close-ups of the areas marked by red and blue squares. References [1] W. Dierking and H. Skriver, “What is Gained by Using an L-band SAR for Sea Ice Monitoring? (84),” in Envisat & ERS Symposium, vol. 572, 2005. [2] W. Dierking, “Synergistic Use of L-and C-Band SAR Satellites for Sea Ice Monitoring,” in 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS. IEEE, 2021, pp. 877–880. [3] L. Eriksson, D. Demchev, A. Hildeman, and W. Dierking, “Alignment of L-and C-Band SAR Images for Enhanced Observations of Sea Ice,” in IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2022, pp. 3798–3801. [4] J. Lohse, A. P. Doulgeris, and W. Dierking, “Mapping sea-ice types from Sentinel-1 considering the surface-type dependent effect of incidence angle,” Annals of Glaciology, vol. 61, no. 83, pp. 260–270, 2020. [5] A. P. Doulgeris, “An Automatic U-Distribution and Markov Random Field Segmentation Algorithm for PolSAR Images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 53, no. 4, pp. 1819–1827, 2015. Towards Multitemporal Sea Ice Classification By Means Of Spaceborne SAR Image Time Series Anja Frost1, James Imber1, Dmitrii Murashkin1, Karl Kortum1, Daniel Gregorek2 1DLR (German Aerospace Center), Remote Sensing Technology Institute, Maritime Safety and Security Lab Bremen, Germany 2MARUM, Center for Marine Environmental Sciences, Bremen, Germany The greatest threat to the Arctic is climate change. Nowhere else is the earth warming up faster than here: Within the last 100 years, the average annual temperature here has risen by five degrees Celsius. Permafrost is thawing. Sea ice is disappearing, more precisely; Multiyear ice has been replaced by seasonal ice, and ice-covered areas have been partially replaced by open water [1]. Therefore, there is a high demand to monitor the Arctic and its sea ice, not only by extend but also by age. Synthetic Aperture Radar (SAR) satellites are able to observe smalland large-scale structures in the sea ice in any weather, through clouds and darkness, due to their active Radar antenna. This makes them well suited to monitor changings in remote areas of the Arctic. However, to analyze large amounts of data sufficiently, automatic algorithms for classifying the sea ice in terms of their stages of development are requested. The development of these algorithms has long been a focus of research ever since the advent of satellite borne SAR [2]. Since then, an abundance of approaches for SAR based sea ice classification have been developed, summarized comprehensively in [3]. The most promising approaches have been transferred to operational services. Nevertheless, obtaining accurate classifications year-round is still a challenge. Different ice classes can show similar radar backscatter responses, which limits the performance of sea ice classification. Seasonally, the radar backscatter signal can be affected by precipitations e.g. wet snow obscures information about underlying ice types [4]. In order to stabilize automated classification, we show here the first tests on multitemporal sea ice classification. That is, we use collocated, sequential SAR acquisitions taken over a region of interest, and – as sea ice is driven by winds and ocean currents – run our sea ice drift retrieval algorithm, which we presented in [5, 6] and improvements of it, for capturing fine-scale sea ice motions especially at the borders of different ice sheets in [7]. Using the retrieved drift vector information, we are able to track drifting pieces of ice (such as an ice floe) from one SAR acquisition to the next, and collect more SAR measurements about the floe. The collected SAR measurements are then used jointly to classify the sea ice. In the study presented here, we only use pairs of SAR images. Larger time series will be considered in future work. We present preliminary test results with Sentinel-1 (S1) data taken over the Arctic Ocean offshore close to Cape Morris Jesup between the Lincoln and Wandel Seas during December 2021, when the ocean showed a closed cover of drift ice (according to dmi ice charts), which most probably consisted of mainly multiyear ice [8-10] and some small leads of open water. East of Cape Morris Jesup, the concentration of multiyear ice may have decreased. Sea ice classification The core of the sea ice classification algorithm is an adjusted UNET++ convolutional neural network architecture described by [11]. In our specific implementation, the classification is done tile-wise, i.e. a S1 acquisition is divided into tiles, classified, and then the results are joined back to generate an ice map as shown in [12]. We differentiate six sea ice types: multiyear ice (MYI), first-year ice (FYI), young ice (YI), open leads (split in so-called dark leads (DL) and bright leads (BL)) and rough ice (RI). For each ice type (and each pixel) a discrete probability distribution over ice type is output. In general, the ice type with the highest probability is then selected for the final ice map. As an example, Fig. 1 shows sea ice classifications (most likely ice type) of two S1 acquisitions taken on 6th Dec. 2021 11:25 UTC and 7th Dec. 2021 17:01 UTC over the study area. Although the sea ice has not changed significantly in the two days according to visual inspection of the radar backscatter signal, both classification results differ a lot. The sea ice classification for the 6th Dec. shows young ice in large areas. The result for the 7th Dec. appears to better reflect the real sea ice situation offshore. Fig. 1: Examples of sea ice classification individually performed on two subsequent Sentinel-1 (S1) acquisitions taken on 6th Dec. 2021 11:25 UTC (left) and 7th Dec. 2021 17:01 UTC (right) over the Arctic Ocean offshore close to Cape Morris Jesup, which is mainly covered with MYI. Even though sea ice has not changed significantly, the classifications differs a lot. Fusion We fuse the discrete probability distributions of the subsequent acquisitions considering the underlying drift. Fig. 2 shows the sea ice drift vector field used for the drift compensation. In the given case, the sea ice moved quite homogeneously eastwards with up to 500 m/h. Fig. 3 is generated out of the two individual classifications, more precisely, the most likely class from both probability distributions is selected after drift compensation. The unlikely presence of “young ice” over Lincoln Sea on 6th Dec. gets automatically corrected in most parts as there is a higher probability of "multi-year ice" the following day. This example showcases the strength of multitemporal sea ice classification. It provides the basis to overcome misclassifications and overall to generate sea ice classifications with increased reliability. Sea ice class MYI FYI YI DL BL RI Fig. 2: Sea ice drift estimated from the two S1 acquisitions used for the sea ice classifications in Fig. 1. Grey scales show S1 HH data from 7th Dec. 2021. Overlaid colours represent retrieved sea ice drift velocity in 500 m resolution and arrows illustrate sea ice drift vectors (10 km spacing). Fig. 3: Sea ice classification based on the two S1 acquisitions used for the sea ice classifications in Fig. 1. The unlikely presence of “young ice” over Lincoln Sea on 6th Dec. gets corrected in most parts. Colour legend see Fig. 1 top right. Discussion For fusing probabilities, various approaches can be applied, namely Kalman filter, Bayesian networks, and Dempster-Shafer. In ongoing work, we consider Kalman filtering and incorporate a priori knowledge to forbid impossible class changes e.g. from “young ice” to “multiyear ice” and vice versa. References [1] Planck, C., Perovich, D. K., & Light, B. (2017, December). Summertime heat and Arctic ice retreat: the role of solar heat on bottom melting of Arctic sea ice. In AGU Fall Meeting Abstracts. [2] Ressel, R., Frost, A., & Lehner, S. (2015). Navigation assistance for ice-infested waters through automatic iceberg detection and ice classification based on TerraSAR-X imagery. The International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, 40(7), 1049. [3] Zakhvatkina, N., Smirnov, V., & Bychkova, I. (2019). Satellite SAR data-based sea ice classification: An overview. Geosciences, 9(4), 152. [4] Kortum, K., Singha, S., & Spreen, G. (2022). Robust Multiseasonal Ice Classification From High-Resolution X-Band SAR. IEEE Transactions on Geoscience and Remote Sensing, 60, 1-12. [5] Frost, A., Jacobsen, S., & Singha, S. (2017, July). High resolution sea ice drift estimation using combined TerraSAR-X and RADARSAT-2 data: First tests. In 2017 IEEE International Geoscience and Remote Sensing Symposium (IGARSS) (pp. 342-345). [6] Frost, A., Wiehle, S., Singha, S., & Krause, D. (2018, July). Sea Ice Motion Tracking from Near Real Time Sar Data Acquired During Antarctic Circumnavigation Expedition. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium (pp. 2338-2341). IEEE. [7] Frost, A., Kortum, K. Wiehle, S., & Tings, B. Ship Navigation Assistance For Polar Waters By Providing Information On Sea Ice Drift And Deformation Zones Using TerraSAR-X Data. Submitted to SeaSAR 2023. [8] Multiyear ice concentration data from 6th and 7th Dec. 2021 are from https://www.meereisportal.de (Förderung: REKLIM2013-04) [9] Ye, Y.; Shokr, M.; Heygster, G. and Spreen, G. (2016), Improving multiyear ice concentration estimates with ice drift, Remote Sensing, 8(5), 397, doi:10.3390/rs8050397. [10] Ye, Y.; Heygster, G. and Shokr, M. (2016), Improving multiyear ice concentration estimates with air temperatures, IEEE Transactions on Geoscience and Remote Sensing, 54(5), 2602–2614, doi:10.1109/TGRS.2015.2503884. [11] Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: Redesigning skip connections to exploit multiscale features in image segmentation,” IEEE transactions on medical imaging, 39(6), 1856-1867. 2019. [12] D. Murashkin, and A. Frost, “Arctic Sea Ice Mapping Using Sentinel-1 SAR Scenes with a Convolutional Neural Network,” IEEE International Geoscience and Remote Sensing Symposium, pp. 5660-5663. 2021. Sea ice drift velocity 0 125 251 376 502 627 759 m/h 1 The impact of input features in deep learning based sea ice mapping Qiang Wang, Malin Johansson, Johannes Lohse, Anthony P. Doulgeris and Torbjørn Eltoft Abstract Here we present a semantic based segmentation model for separating sea ice from open water in Sentinel-1 imagery. The study utilizes Sentinel-1 intensity imagery, together with sea surface temperature (SST) obtained from Sentinel3 data, to train a UNET based semantic segmentation model for separating sea ice from open water. The imagery were all acquired over an area between Greenland and Svalbard, and contains images from all seasons. Two network configurations were studied: A ”baseline model”, which used the HH, HV intensities and the incident angle, and an ”advanced model”, which in addition included the SST as inputs. We compared the models’ performances on independent Sentinel-1 scenes acquired over different seasons. The results indicate that both the baseline and the advanced models are able to separate the open water and sea ice classes well. Improvements were identified in the delineation of the sea ice edge for the advanced model. I. Introduction Monitoring sea ice and its changes in extent and concentration has drawn more attention with the ongoing global warming [1]. Wunderling et al [1] indicated that the Arctic Ocean might become ice-free during summer within the 21st century. Meanwhile, Post et al [2] reported that sea ice loss could influence terrestrial productivity and diversity, species interactions, etc. Hence, a better understanding of temporal and spatial changes in sea ice is important for improving climate modelling, understanding the Arctic ecology, as well as supporting safe marine navigation and offshore operations. Ship-based expeditions are a good way to observe sea ice conditions with high accuracy, though are expensive and time consuming and hence not feasible to be carried out regularly over large spatial scales. Remote sensing can provide data for regular monitoring. For example, Spreen et al [3] developed a sea ice concentration algorithm called ARTIST (ASI) using data from Advanced Microwave Scanning Radiometer (AMSR-E) and Advanced Microwave Scanning Radiometer 2 (AMSR2). The maps are openly available, regularly updated and has a km resolution. However, such relatively coarse spatial resolution might not be sufficient for accurate marine navigation. Synthetic aperture radar (SAR) sensors, e.g. Sentinel-1, on the other hand, can provide measurements at much finer spatial resolution on the order of tens of meters resolution. Park et al [4] developed a machine learning sea ice type classifier by using the texture features from Sentinel1 and retrieved three generalized ice types (open water, mixed first-year ice, old ice) with an overall accuracy of 87% (winter)and 67% (summer). Wang and Li et al [5] developed an approach for deriving a high-resolution sea ice product for the Arctic Ocean by employing an integrated stacking model to combine multiple UNET [6] classifiers with diverse specializations with Sentinel-1 imagery. It obtained an overall accuracy of 96.10% by using the HV-polarization, polarization ratio and polarization difference as input features. Here we propose in this paper two simplified UNET configurations for sea ice and open water separation. One model, denoted the ”baseline model”, uses the HH, HV intensities and the incident angle (IA) as input features, whereas the second model, denoted the ”advanced model”, also includes sea surface temperature (SST) as input. The networks were trained using Sentinel-1 (baseline), and SST from Sentinel-3 (advanced), and their performances were subsequently validated with regard to accuracy and robustness. II. Data and methodology A. Data We utilized Sentinel-1 Ground Range Detected (GRD) in the Extra Wide (EW) swath mode over the area between Greenland and Svalbard. Sentinel-1A + 1B offers a 6 day exact repeat cycle. At medium resolution, the GRD product is provided at a spatial resolution of 40 ×40 m and each scene covers roughly 10,000 ×10,000 pixels. In total, 32 Sentinel1 images acquired during all the months of 2021 were used to generate the training samples for our experiment. Scenes in both ascending and descending mode were used in order to prevent biased model fitting. Daily SST was obtained from the Copernicus climate change service with the spatial resolution of 0.05° × 0.05°, and were re-sampled to match the spatial resolution of the Sentinel-1 images. INVESTIGATION OF MULTIFREQUENCY SAR IMAGE ALIGNMENT BY ICE DRIFT COMPENSATION IN THE MARGINAL ICE ZONE Denis Demchev, Leif E. B. Eriksson, Anders Hildeman∗ Chalmers University of Technology Gothenburg, Sweden ∗A.H. is currently with the AstraZeneca AB Wolfgang Dierking The Arctic University of Norway Tromso, Norway Alfred Wegener Institute Helmholtz Centre for Polar and Marine Research Bremerhaven, Germany ABSTRACT A recently proposed SAR image alignment framework based on ice drift compensation is examined for applicability in the marginal ice zone where ice analysis is complicated by ice floe rotations, fast changing local ice concentration, local wind variations, and changing wetness of the ice surface. Here, we use L-band ALOS-2 and C-band Sentinel-1 images that were acquired over the Greenland Sea in the summer period of 2019. Preliminary results show that L-C SAR image alignment is possible in the marginal ice zone during melting conditions with a time span of a few hours and with ice speeds of >10 cm/s. The main challenges and directions of future work to improve the alignment performance in the transitional zone between open sea and dense drift ice are formulated. Index Terms—Alignment, sea ice drift, marginal ice zone, SAR, Arctic 1. INTRODUCTION The marginal ice zone (MIZ) denotes the transition between the open sea and dense drift ice or fast ice and is of large interest for research on sea dynamics as it is characterized by various features comprising ice edge dispersion/compacting, formation of ice bands, intense air-sea interaction which presence of oceanic eddies [1]. Satellite data from synthetic aperture radar (SAR) are extremely valuable for sea ice monitoring including the MIZ. However, the automated retrieval of ice movements in the MIZ from pairs of SAR images can be complicated because of ice floe rotations, wind-induced changes of the water surface roughness, and wetness of the ice surface. The usage of SAR images acquired at different frequencies is beneficial as it provides richer information for monitoring and retrieval of sea ice parameters. Most often the images at different frequencies are not acquired from the same satellite but rather from different satellites observing the same region at slightly different points in time. This poses a problem for their joint analysis as ice drifts due to wind and sea currents will transform the ice in-between the two paired acquisitions. To collocate the SAR sea ice images, a recently proposed SAR ice imagery alignment method based on ice drift compensation [2] is used to assess the feasibility of alignment in the marginal ice zone. 2. SAR IMAGES AND TEST SITE We collected L-C image pairs from Level-1.5 Ground Range Detected (GRD) data from ALOS-2 (Advanced Land Observing Satellite-2) and Sentinel-1A/B GRD Extra Wide Swath Mode (EW) data taken over the ice edge area in the Greenland Sea. Two different ALOS-2 data modes are used, namely ScanSAR Nominal and Stripmap Fine. The backscattering coefficient σ0was computed and the data were projected onto a Polar Stereographic grid with a pixel size of 100 m. 3. SAR ALIGNMENT METHOD AND QUALITY ASSESSMENT Alignment of SAR images acquired at Land C-band was demonstrated in [2]. This study examines the performance of the alignment method in the marginal ice zone. An L-C pair has been selected in the Greenland Sea area (10.8191◦W, 79.2014◦N) in the summer period (2019/06/22) with a time gap between acquisitions of around 4 hours (Figure 1 A, B). A SAR drift algorithm [3] is used to estimate the ice drift for the overlapping parts of the SAR images, which provides the ice displacement vectors on a regular grid with a step size of 50 pixels, corresponding to 5 km. The alignment is carried out by local piece-wise affine image transforms on a triangular mesh obtained based on the ice displacements. The alignment framework was demonstrated with pairs of ALOS-2 and Sentinel-1 images, but can be used for data at other frequencies and from other sensors. For the alignment quality evaluation, we use the Structural Similarity Index (SSIM) [4] which is a similarity measure comprising brightness, contrast, and structural components. Its values range from -1 to 1, with the latter indicating a perfect alignment. For consistency in the assessment of alignment quality, we apply SSIM for the original L-band image L, and image LF B which is obtained after forward transform TF followed by backward transform TB:LF B =TB(TF(L)). But it should be noted that we use the two-times transformation as some kind of proxy for assessing one-time transformation to exclude the influence of the signature difference. Fig. 1. A - L-band ALOS-2 HH image from 2019/06/22 13:45:11; B - C-band Sentinel-1 HH image from 2019/06/22 17:40:12; C - aligned ALOS-2 image by ice drift compensation. The derived ice displacement vectors are depicted in green on the Sentinel-1 image (B). 4. RESULTS AND DISCUSSION Ice drift retrieval in the marginal ice zone becomes more complex compared to pack ice mainly due to two factors: (1) more intense dynamics of ice floes including their rotations and (2) the influence of open water patches that constrains the ice signature recognition. As the drift algorithm [3] handles the rotational movement of ice, we obtained displacement vectors for almost each grid point, which have been visually checked for correctness. According to the drift algorithm output, the mean ice speed in the period was 14.7 cm/s with a dominating drift towards NE and N. Individual ice floes with a horizontal size from a few hundred to tens of kilometers can be recognized in Figure 1. A result of the alignment of the ALOS-2 image to the Sentinel-1 image by ice drift compensation is shown in Figure 1 C. Visually, it is seen that the aligned ALOS-2 image has not been significantly distorted after alignment and compared to the original image (Figure 1 A). This indicates the correctness of the obtained drift vectors and successful transformation by the alignment method. To assess the quality of the alignment quantitatively, we used the SSIM values. Here we assess the transformation quality but not the differences caused by sea ice signature changes at the different bands, therefore we applied the metric for the original ALOS-2 image and its forward-backward transformed version. Figure 2 shows that the SSIM values close to 1 are prevailing across much of the aligned image, while the lower values correspond to a relatively small fraction of pixels. The mean air temperatures over the area between 13:00 and 18:00 on 2019/06/22 were derived from the ERA5 reanalysis dataset [5] and are depicted in Figure 2. According to the data, the prevailing air temperature at 2 m over the entire region was above 0◦C in previous days, and in the period between image acquisitions, hence we can expect melting conditions. 5. CONCLUSIONS In this case study, we assessed the applicability of the recently proposed alignment framework for SAR sea ice imagery in presence of granular ice covers consisting of relatively small, thin ice floes, which are common in the marginal ice zone. Such analysis is more challenging compared to pack ice due to the complex nature of the MIZ as described above. Successful alignment has been demonstrated for a pair of L-band ALOS-2 and C-band Sentinel-1 images from 2019/06/22 acquired over the Greenland Sea with a difference in acquisition time of 4 hours. The feasibility of alignment has been examined visually by the presence of distortion as well as quantitatively by the Structural Similarity Index. Both the .visual assessment and the SSIM confirm the applicability and efficiency of the alignment method based on the piecewise affine transform in summer conditions, rotational ice movement, and locally varying ice concentrations. Despite promising results, it has to be proven that the alignment can be applied for time gaps longer than a few hours and under different weather conditions. The ice drift algorithm used in this study has demonstrated reliable performance for multifrequency SAR images over the marginal ice zone, but the image alignment can also be assessed for other SAR drift algorithms. 6. ACKNOWLEDGEMENT This work was done as part of the European Space Agency contract no. 4000130509/20/NL/FF/ab, ”Synergistic Use of Land C-band SAR Satellites for Sea Ice Monitoring”. ALOS-2/PALSAR-2 data are provided by JAXA through the 2019 to 2022 mutual cooperation project between ESA and JAXA on Using Synthetic Aperture Radar Satellites in Earth Science and Applications. Sentinel-1 data (European Union Copernicus program) were downloaded from the NASA Alaska Satellite Facility (ASF) SAR Distributed Active Archive Center (DAAC). Fig. 2. A - Structural Similarity Index (SSIM) from the original Land forward-backward transformed image LF B ; B - the mean air temperature at 2m from ERA5 reanalysis between 13:00 and 18:00 on 2019/06/22. 7. REFERENCES [1] Matti Lepp¨ aranta, The drift of sea ice, Springer Science & Business Media, 2011. [2] Leif EB Eriksson, Denis Demchev, Anders Hildeman, and Wolfgang Dierking, “Alignment of l-and c-band sar images for enhanced observations of sea ice,” in IGARSS 2022-2022 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 2022, pp. 3798–3801. [3] Anders Berg and Leif EB Eriksson, “Investigation of a hybrid algorithm for sea ice drift measurements using synthetic aperture radar images,” IEEE transactions on geoscience and remote sensing, vol. 52, no. 8, pp. 5023– 5033, 2013. [4] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004. [5] Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr´ as Hor´ anyi, Joaqu´ ın Mu˜ noz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, et al., “The era5 global reanalysis,” Quarterly Journal of the Royal Meteorological Society, vol. 146, no. 730, pp. 1999–2049, 2020. 1 High resolution Land C-band polarimetric variability during MOSAiC A. Malin Johansson1, S. Singha2, G. Spreen 3, S. Howell 1 1UiT The Arctic University of Norway 2National Center for Climate Research (NCKF), Danish Meteorological Institute (DMI) 3Institute of Environmental Physics, University of Bremen 4Climate Processes Section, Science and Technology Branch, Environment and Climate Change Canada / Government of Canada Abstract The polarization difference (VV-HH) for different sea ice types observed in Land C-band SAR images during the MOSAiC drift study is investigated here. The drift lasted from the freeze-up to the early melt season, ensuring that the temperature dependency is also examined. PD has positive values for open water and new ice areas, i.e. HH >VV, whereas the values turn negative for the young ice stage and stabilise around 0 for the thicker sea ice types. L-band SAR PD values appeared to have a higher sensitivity to the melt onset compared to C-band SAR. I. Introduction In September 2019, the German research icebreaker R/V Polarstern started the drift across the Arctic Ocean as a part of the MOSAiC (Multidisciplinary drifting Observatory for the Study of Arctic Climate) drift experiment. During the course of MOSAiC thousands of SAR images were acquired in different imaging mode and frequency supported by various international space agencies and research organizations (e.g. CSA, JAXA, ESA, DLR, KARI, ASI, CONAE, INTA). In this study we focus on SAR images from the ALOS-2 and RADARSAT-2 satellites acquired through joint ESA-JAXA collaboration. Here we evaluate the effects of seasonal changes on Cand L-band backscatter in respect to four different sea ice types, i.e., newly formed sea ice (NI), Young Ice (YI), Smooth Ice (SI) and Deformed Ice (DI), as well as the stability of radar backscatter in different polarization, for polarimetric parameters acquired under different environmental conditions and at different incidence angles. Previous studies e.g., [1–3] have shown that L-band SAR can provide improved separability between different sea ice types. The different wavelengths means e.g., different penetration depths and different sensitives to the onset of melting e.g. [4]. In [2] L-band data was shown to provide easier separation between first year ice (FYI) and second/multi-year ice (MYI) in the early and advanced melt season. Upcoming L-band missions such as NISAR, ALOS-4 and ROSE-L will offer the advantages of fully polarimetric acquisitions along with higher ground coverage to achieve a optimal scenario for L-band SAR based sea ice monitoring. In this study we found that the polarization difference (PD) can be used to separate out younger sea ice types in both Cand L-band SAR images, provides a complement to the co-polarization ratio that can primarily separate the newer sea ice types, and PD provide good separability between high backscatter young ice and MYI. PD could also be used as an indicator of early melt stages, due to the shift from stronger VV to stronger HH channel data in L-band SAR images. The same trend could not be observed in the C-band images, and as such the two frequencies complemented one another. II. Data set SAR data overlapping the drift of R/V Polarstern during the MOSAiC expedition from 1 October 2019 until 31 July 2020. The sea ice observed in the larger area around R/V Polarstern for the first floe is the main focus of this study. R/V Polarstern drifted towards the polar hole (ca. above 86.6oN) around the end of 2019 and the beginning of 2020. Some sensor specifics are listed in Table I. TABLE I Specifics of satellite data used in this study. The values presented are average values for the different missions. Mission Frequency Range x Azimuth Range x Azimuth IA range NESZ resolution* spacing * RADARSAT-2 C (5.41 GHz) 5.2 m x 7.6 m 4.7 m x 5.1 m 18 o–49 o-32.9db ±1.5dB ALOS-2 L (1.2 GHz) 5.1 m x 4.3 m 2.9 m x 3.2 m 29 o–41 o-36.0 dB(HH), -46.0 dB (HV) 2 III. Method Here we analyse the polarimetric SAR information for four different sea ice types, NI, YI, SI and DI, and how they vary with incidence angles, SAR frequencies and temperature. The NI was predominantly observed within leads and the YI was predominantly the younger sea ice surrounding the MOSAiC floe at the start of the campaign. This YI had on 14th October an average thickness of 30 cm. The SAR images were calibrated so that the pixel values could be directly related to the radar backscatter coefficient (σ0) values. The backscatter coefficients are derived from the mean intensity of the complex scattering coefficient S, where ⟨·⟩ is averaging over a neighbourhood of N pixels. Here N is equal to 7x7 pixels. The polarization difference (P D) was calculated as; P D =⟨|SV V |2⟩ − ⟨|SHH |2⟩.(1) Note that P D is defined on a linear scale, and using P D the non-polarized component can be separated from the polarized one in the analysed data, and is less sensitive to the closeness to the NESZ than many other polarimetric features. A poor signal-to-noise ratio (SNR) was in, e.g., [5], shown to result in an over representation of perceived volume scattering when SNR is less than 10 dB. A low SNR primarily affects the NI areas as well as the refrozen melt pond parts of the MOSAiC floe and similar floes in the vicinity. The sea ice growth during the season means that poor SNR is a larger problem at the start of the campaign, with the overall thinner sea ice as well as a larger proportion of refrozen melt pond floes. Moreover, during the melt season stage when R/V Polarstern was located in the marginal ice zone poor SNR was more prominent. Region of interests (ROIs) representing the four different ice types were manually identified. During the selection of these areas we mainly relied on ice charts, ice concentration map and expert visual judgment of the Pauli RGB compositions. The areas were selected in such a way that they adequately represent the incidence angle variation from near to far range. Not all ice types were available within all images, in particular were the NI and YI not always present. When possible was the MOSAiC floe included in the ROIs. IV. Results and Discussion During the The MOSAiC drift expedition the sea ice drift remained reasonably stable from the start until December, and during this time period we also observe stable backscatter values that primarily vary with incidence angle. Once R/V Polarstern drifted out of the polar hole increased overall sea ice deformation could be observed. In all frequency bands the co-pol power ratio has been proven to be an excellent measure to distinguish open water and newly formed sea ice from the thicker sea ice types, SFYI and RFMYI ([1, 6]). In the case of younger ice types, the co-pol power ratio is not as powerful in discriminating thin ice from open water. PD shows some promise in the separability of the NI and YI areas as well as separation from thicker ice types. In this study we focus on changes in backscatter and PD within the SAR images from the freeze-up to the melt onset. For background do open water and NI have high PD values, though once the transition to YI areas occur the PD values are reduced and for thicker sea ice the mean values stabilise around 0. This follows the trend seen at other sea ice campaigns such as N-ICE2015 and Beaufort 2015 during the freeze-up and winter months. In Figure 1 the PD evolutions for the Land C-band images during MOSAiC is shown. For all ice types does L-band data have lower variability than overlapping C-band data. The transition in PD values from the NI to YI is likely dependent on the penetration depth as the transition occurs later in time in the L-band data compared to the C-band data for the sea ice areas surrounding the MOSAiC floe co-inside with the thermodynamic sea ice growth. The L-band images had higher (and positive) PD values at the start of the freeze-up than the C-band images for the thin ice areas (30 cm thick) surrounding the ship and the MOSAiC floe, though these values stabilised once the sea ice had grown in thickness and transitioned to negative values by mid-November. The negative values for both Cand L-band lasted all the time until R/V Polarstern entered the polar hole. Analysis of the scattering mechanisms in the area indicate predominant surface scattering for the thinner sea ice area from the freeze-up to the end of December. Starting with the freeze-up season we can observe that the SI largely have low PD values and a low variability whereas the DI have mean values close to 0 but with larger standard deviation (std). PD overall has a larger variability for deformed sea ice compared to the other ice types for both frequencies, and as such can the coefficient of variation (CV) or std over an area be used to separate the DI from the surrounding sea ice types. This was also confirmed by investigating MYI areas in Canada. For the MOSAiC floe this meant that the refrozen melt pond areas had low PD values with low std, and the deformed part of the floe had high std values. Once Polarstern drifted out of the polar hole the temperatures started to increase and by mid-April they were above -10oC, with temperatures around 0oC in mid-April. A PD temperature dependency was observed once the temperatures remained above 0oC the mean value and the std increased significantly in the L-band data, meaning that this parameter may be possible to use to indicate melt onset. For the C-band data there was an increase in variability 3 Fig. 1. The PD evolution for RADARSAT-2 (top) and ALOS-2 (bottom). for the SI and a decrease for the DI. Combining the PD evolution for the different sea ice types and frequencies makes the two frequencies complementary for sea ice type separation and classification. V. Conclusions The polarization difference was found to provide separability between younger ice types and thicker sea ice and provide good separability between high backscatter young ice and MYI. The variability in PD was in L-band SAR was found to be a useful indicator of early melt stages, due to the shift from stronger VV to stronger HH data. SAR images from the CIRFA cruise in 2022 at the Belgica Bank fast ice was found to confirm this analysis and the usefulness of PD will be further investigated using both the MOSAiC, N-ICE2015 and the CIRFA cruise data, utilizing overlapping Land C-band images that co-inside with in-situ measurements. Acknowledgement The work was funded through the Research Council of Norway Grant no. 237906. RADARSAT-2 Data and Products ©MDA Geospatial Services Inc. 2019/2020. RADARSAT is an official mark of the Canadian Space Agency References [1] A. M. Johansson, C. Brekke, G. Spreen, and J. A. King, “X-, C-, and L-band SAR signatures of newly formed sea ice in Arctic leads during winter and spring,” Remote Sensing of Environment, vol. 204, pp. 162 – 180, 2018. [2] M. Mahmud, V. Nandan, S. Howell, T. Geldsetzer, and J. Yackel, “Seasonal evolution of l-band sar backscatter over landfast arctic sea ice,” Remote Sensing of Environment, vol. 251, p. 112049, 2020. [3] T. Toyota, J. Ishiyama, and N. Kimura, “Measuring deformed sea ice in seasonal ice zones using l-band sar images,” IEEE Trans. Geosci. Remote Sens., vol. 59, no. 11, pp. 9361–9381, 2020. [4] J. A. Casey, S. E. Howell, A. Tivy, and C. Haas, “Separability of sea ice types from wide swath Cand L-band synthetic aperture radar imagery acquired during the melt season,” Remote Sensing of Environment, vol. 174, pp. 314 – 328, 2016. [5] M. Espeseth, C. Brekke, C. Jones, and A. Holt, B. abd Freeman, “The impact of system noise in polarimetric sar imagery on oil spill observations,” IEEE Trans. Geosci. Remote Sens., vol. 58, no. 6, pp. 4194–4214, 2020. [6] T. Geldsetzer and J. J. Yackel, “Sea ice type and open water discrimination using dual co-polarized C-band SAR,” Canadian Journal of Remote Sensing, vol. 35, no. 1, pp. 73–84, 2009. 1 Tracking Backscatter Signatures of Individual Sea Ice Floes Using In-Situ Ice Drift Observations Catherine Taelman1, Johannes Lohse1, Anthony P. Doulgeris1 1) UiT The Arctic University of Norway Abstract Studies on seasonal variations in radar backscatter exist for landfast ice, but seem sparse for drifting sea ice. Here, we introduce a new in-situ sea ice drift data set collected by 17 drifters deployed during the CIRFA-22 cruise in Fram Strait in April-May 2022. The data set covers the transition from freezing conditions to melt onset in Fram Strait and the Greenland Sea. These drift trajectories provide the rare opportunity to track individual sea ice floes and study the temporal evolution and incident angle dependence of their radar backscatter signatures as they undergo physical changes during melt onset. We furthermore demonstrate how the amount of tracked floes can be increased by identifying distinct surface structures in the vicinity of the drifter location in consecutive radar images. I. Introduction Because of its fine spatial resolution and its independence of daylight and cloud conditions, synthetic aperture radar (SAR) is a primary tool for operational sea ice monitoring. The appearance of different sea ice types in SAR imagery is determined by radar parameters, such as polarization, frequency, and incident angle (IA), and surface parameters, such as small-scale roughness, large-scale deformation, and the dielectric properties of the ice and snow layers. At present, most of the SAR data that is routinely available for sea ice monitoring is acquired in wide-swath mode at C-band, for example by Sentinel-1 (S1) or Radarsat-2 (RS2). For these sensors, the backscatter signal from a given sea ice type varies with IA across the swath [1], and it is strongly affected by changes in temperature and snow properties [2]. Understanding these variations in backscatter is crucial for the interpretation and automated analysis of the imagery. The effect of melt onset and the seasonal evolution of C-band radar backscatter over the yearly cycle has been extensively studied for landfast ice (for example [3, 4]). However, studies tracking the backscatter signature of individual drifting sea ice floes seem sparse. Here, we use a set of in-situ drift trajectories to track individual ice floes in the Greenland Sea over several months, covering the transition from freezing temperatures to warmer melting conditions. We automatically identify the individual drifter locations in overlapping S1 imagery and extract backscatter intensities from the pixels around the drifter. Furthermore, we visually identify and track the backscatter of distinct surface features in the vicinity of the drifters that can be followed over time. On short timescales (days), this allows us to investigate the IA dependence of the exact same sea ice, while on longer timescales (weeks), we can study the backscatter evolution of drifting ice floes with changes in temperature during melt onset. The main focus of this work can be summarized in three points: •We introduce a unique data set of in-situ drift observations in the Greenland Sea in the time period from April until December 2022. •We demonstrate the use of the drifter trajectories to track individual sea ice floes and investigate the IA dependence and the temporal evolution of their backscatter signature. •We conceptually show how to expand the tracked ice areas by identifying distinct surface structures that are clearly visible in the vicinity of the drifter locations. II. Data Set A. In-situ sea ice drift data We use in-situ sea ice drift observations collected from drifters that were deployed during the CIRFA-22 cruise to Fram Strait and Belgica Bank in the time period from April 22nd to May 10th 2022. A total of 17 sea ice drifters were manufactured in-house at UiT after the design of Rabault et. al. [5]. They sample a GPS position at a 30 minutes interval and measure wave motion for a duration of 20 minutes every 3 hours. Of the 17 sea ice drifters, 15 were deployed manually on the sea ice, while 2 were set out by a drone. At the sites of manual deployment, additional in-situ measurements were acquired: ice thickness, snow depth, snow-water equivalent (SWE), and snow salinity at the snow-ice interface. The floes tagged with drifters continuously move southwards into the Greenland Sea, where they eventually break up or melt. The sea ice drifters then end up in the open water and transition into ocean drifters. They float in the water and continue sampling location and wave motion. By analyzing the wave motion measurements, we can determine when this ice-to-water transition takes place (see Section III). 2 B. Sentinel-1 data We use freely available Sentinel-1 (S1) imagery to study the temporal evolution of SAR backscatter from the sea ice at the drifter deployment sites. S1 operates at C-band and is at present one of the main data sources for operational sea ice monitoring. All S1 data used in this study was acquired in dual-polarimetric extra-wide swath (EW) mode and processed to ground range detected format at medium resolution (GRDM). The EW GRDM product is provided at 40x40 m pixel spacing at a swath width of 410 km. Fig. 1. Left: Full drift trajectories of all 17 drifters that were deployed during the CIRFA-22 cruise. Right: Partial trajectory of the example drifter (ID 206754) used in this abstract. The markers indicate the drifter location at the time of the overlapping S1 images shown in Figure 2. III. Method A. Drifter data processing A full drifter trajectory consists of a sea ice part, followed by an ocean part. By careful analysis of the wave spectra time series, which are obtained from the wave motion measurements, we identify 3 stages in each trajectory: sea ice - transition - ocean. In the sea ice stage, there is little to no wave motion, and the drifter is clearly on sea ice. During the transition phase, the floe with the drifter starts breaking up or melting as it reaches the marginal ice zone, which is reflected by an increase in wave motion. Finally, the drifter ends up in the open water (ocean). This stage is characterized by a major rise in wave motion. We flag all drifter trajectories according to these three stages, and only use data from stage 1 (sea ice) for this study. B. Identifying overlapping Sentinel-1 data To find S1 scenes that overlap with the drifter positions, we loop through the time series of each individual drifter as long as it is on sea ice. For every data point, we use the SentinelAPI from the sentinelsat python package to search for overlapping S1 images. As the drifters sample a GPS location every 30 minutes, we search over a time interval from 15 minutes before until 15 minutes after the drifter location timestamp. This results in a maximum time difference of 15 minutes between the S1 image acquisition and the recorded drifter location. Note that the ice floe can drift a distance that corresponds to several image pixels during this time. Assuming a drift speed of 0.1 m s, the covered distance will be 90 m (that is 2.25 pixels) in 15 minutes. C. Extracting backscatter intensities The drifters enable us to track backscatter intensities in two different ways. 1) The most straightforward approach is to extract a patch of pixels directly around the location of the drifter, and compare these patches throughout the time series. This more conventional way of tracking sea ice is, however, limited when the drifter is close to the edge of the floe, or when the sea ice deforms or breaks up in the direct vicinity of the drifter. Moreover, the drifter locations in the S1 imagery need to be drift-corrected if the time difference between image acquisition and GPS timestamp is large enough to result in a significant location offset in the SAR image. 2) We expand this conventional method by tracking distinct surface structures that are clearly recognizable in the SAR image in the area close to the drifter location. As a proof of concept, we identify these structures manually. At a later stage, this step can be automated using feature tracking and pattern matching algorithms. We investigate the IA dependence by comparing backscatter values of one floe imaged in two consecutive S1 scenes, once in near-range and once in far-range. Using both proposed workflows to track sea ice, we study the temporal evolution of the backscatter signatures of distinct sea ice structures and floes. 3 IV. Preliminary Results We collected a unique data set of in-situ sea ice drift observations in the Greenland Sea during the transition from freezing to melting conditions in 2022. The trajectories of the 17 drifters are shown in Figure 1. Note that these trajectories include the entire time series (sea ice - transition - ocean), ranging from April to December 2022. (see Section II-A). A small selection of the overlapping S1 image time series for one example drifter is shown in Figure 2 (top). The drifter location is marked with a red star. We identified two distinct sea ice structures throughout the time series and selected a region of interest (ROI) in each of them, visualized as purple (ROI 1) and blue (ROI 2) rectangles. Note that ROI 2 is not drawn in the S1 image of 26/05/22, as it is not clearly recognizable there. Scatterplots of the Sigma Nought backscatter values in HV versus HH polarization for a small area around the drifter location and for both ROI’s at the timestamps of the S1 images is shown in Figure 2 (bottom). Fig. 2. Top: False-color RGB images (R-HV, G-HH, B-HH) of four S1 scenes that overlap with the drift trajectory of drifter 206754. The images are cropped to areas of 48x48 km around the drifter location at the time of image acquisition, as indicated in Figure 1 (right). The drifter location is indicated with a red star. Purple and blue rectangles indicate ROI’s on distinct sea ice structures. Bottom: Scatterplots of HH versus HV backscatter values for a small area around the drifter location and for the two ROI’s at the timestamps of the selected S1 images. Future work includes tracking the backscatter signatures over the full temporal time series for this case study and for more floes and distinct surface structures, using the CIRFA-22 drift data set. The larger sample size will allow for a representative analysis of the temporal variations and incidence angle dependence of radar backscatter during melt onset. References [1] J. Lohse, A. P. Doulgeris, and W. Dierking, “Mapping sea-ice types from Sentinel-1 considering the surface-type dependent effect of incidence angle,” Annals of Glaciology, vol. 61, no. 83, pp. 260–270, 2020. [2] D. Barber, E. LeDrew, D. Flett, M. Shokr, and J. Falkingham, “Seasonal and diurnal variations in SAR signatures of landfast sea ice,” IEEE transactions on geoscience and remote sensing, vol. 30, no. 3, pp. 638–642, 1992. [3] J. Yackel and D. Barber, “Melt ponds on sea ice in the Canadian Archipelago: 2. On the use of RADARSAT-1 synthetic aperture radar for geophysical inversion,” Journal of Geophysical Research: Oceans, vol. 105, no. C9, pp. 22 061–22 070, 2000. [4] R. Scharien, J. Yackel, M. Granskog, and B. Else, “Coincident high resolution optical-SAR image analysis for surface albedo estimation of first-year sea ice during summer melt,” Remote sensing of environment, vol. 111, no. 2-3, pp. 160–171, 2007. [5] J. Rabault, T. Nose, G. Hope, M. M¨uller, Ø. Breivik, J. Voermans, L. R. Hole, P. Bohlinger, T. Waseda, T. Kodaira et al., “Openmetbuoy-v2021: An easy-to-build, affordable, customizable, open-source instrument for oceanographic measurements of drift and waves in sea ice and the open ocean,” Geosciences, vol. 12, no. 3, p. 110, 2022. Sea ice cover and drift by Sentinel-1 SAR and the support for Arctic shipping Xiao-Ming Li Aerospace Information Research Institute, Chinese Academy of Sciences Widely used sea ice cover and drift data in polar regions are derived mainly from spaceborne microwave radiometer and scatterometer data, and the typical spatial resolution of these products ranges from several to dozens of kilometers. Due to dramatic changes in polar sea ice, highresolution sea ice data are drawing increasing attention for polar navigation, environmental research, and offshore operations. In this paper, we focused on developing methods for deriving a high-resolution sea ice volume and dynamics data, i.e. sea ice cover and drift using Sentinel-1 (S1) SAR dual-polarization data in extra wide swath (EW) mode. The approach (Fig.1) for discriminating sea ice from open water by S1 HH and HV (denoised, as examples shown in Fig.2) data is based on a modified U-Net architecture (Fig.3), a deep learning network. By employing an integrated stacking model to combine multiple U-Net classifiers with diverse specializations, sea ice segmentation is achieved with superior accuracy over any individual classifier (example shown in Fig.4). We applied the proposed approach to over 28,000 S1 EW 15 images acquired in 2019 to obtain sea ice cover products in a high spatial resolution of 400 m. By converting the S1-derived sea ice cover to concentration and then compared with Advanced Microwave Scanning Radiometer 2 (AMSR2) sea ice concentration data, showing an average absolute difference of 5.55 % with seasonal fluctuations (Fig.5). A direct comparison with Interactive Multisensor Snow and Ice Mapping System (IMS) daily sea ice cover data achieves an average accuracy of 93.98 %. These results show that the developed S1-derived sea ice cover results are comparable to the AMSR and IMS data in terms of overall accuracy but superior to these data in presenting detailed sea ice cover information, particularly in the marginal ice zone. So far we have generated the S1-derived sea ice cover dataset from 2019-2022. Fig.2 Flowchart of the proposed method for deriving sea ice cover information from S1 EW images in HV and HH polarization.  Fig.1 Example of the process of S1 data: (a) original HH-polarized data (b) incidence anglecorrected HH-polarized data (c)original HVpolarized data and (d)denoised HV-polarized data using the method proposed by Sun and Li (2020). Potential Application of the Earth Explorer 10 candidate Harmony for Sea Ice Model Validation Sea ice in the Arctic Ocean is in continuous motion under influence of wind, ocean, and internal stress. Sea ice deformation (computed as a spatial derivative of the drift field) is localized in space and time and forms elongated, narrow zones, also called Linear Kinematic Features (LKFs). The frequency distribution of deformation rates as well as the pattern, density, orientation, and intersection angle of LKFs are a characteristic feature of sea ice. Sea ice drift and deformation can be observed by pattern matching techniques applied to passive microwave, or scatterometer, or synthetic aperture radar (SAR) satellite data. The disadvantage of the type of ice drift product is the large time delta required between the image acquisitions, and a relatively low spatial resolution. Harmony is a candidate for the Earth Explore 10 mission. The two Harmony satellites will fly in a reconfigurable formation with Sentinel-1D. Both will be equipped with a multi-angle thermal infrared sensor and a passive radar receiver, which receives the reflected Sentinel-1D signals using two antennas. In the stereo formation, the Harmony satellites will fly approximately 300 km in front and behind Sentinel-1, which allows for the estimation of instantaneous sea-ice drift vectors. As it was shown by Kleinherenbrink et al., [2021] the sea ice drift and deformation can be derived from simulated Harmony data, but the signal-to-noise ratio is quite low. The goal of this study is to evaluate the applicability of the Harmony data for statistical characterisation of sea ice deformation in the Arctic Ocean and feasibility in utilisation for tuning parameters of the next generation sea ice model (neXtSIM, [Olason et al., 2022]). NeXtSIM can realistically simulate the sea ice motion and deformation. Both the LKFs and the spatial scaling of the sea ice deformation rate simulated by neXtSIM compare quite well to PMW and SAR observations. Sensitivity of synthetic Harmony data to neXtSIM rheological parameters is studied. A scenario of model validation against instantaneous sea ice drift satellite estimates is suggested. Denoising Doppler shift data Computing the Doppler shift signal and adding thermal noise is performed with the same forward model (FM) as in [Kleinherenbrink et al. 2021]: D, ND = FM (U) This yields 2D fields of Doppler shift from Concordia, Discordia, and Sentinel (DC(R,A), DD(R,A), DS(R,A)) and the thermal noise equivalent sigma zero (NESZ) of the Doppler signal (ND(R,A)). Each Doppler shift field is further treated separately and the subscripts and (R,A) dependence is omitted for clarity. Thermal noise is added as a product of NESC profile and normally distributed noise (N): DN = ND * N Noise correction is further performed for each field of D individually. Given that the profile of NESZ is known a priori we can perform “texture noise” correction suggested by Park et al. [2019] for reducing amplitude of signal variations near inter-swath boundaries where NESZ is the highest: D1 = Gf(DN)*NN + DN*(NN – 1), Where Gf is a 2D Gaussian filter with size of 10 pixels (20 km) and NN is ND normalized into range 0 – NMAX, with NMAX = 0.7 being found empirically. Next, the low-pass filter is applied to D1 as suggested in [Kleinherenbrink et al. 2021, eqs. 19 – 22]: D2 = Kf(D1, D). Note, that the original fields with Doppler shift (D, prior to adding noise) are required to compute the cut-off frequency. Finally, the anisotropic diffusion filter [Perona and Malik, 1990] is applied is applied for smoothing homogeneous U/V fields and preserving high contrasts: D3 = ADf(D2, gamma=0.25, kappa=5). Denoising instantaneous sea ice drift velocities Velocity fields (U) are reconstructed from D, DN and from D3 using the same retrieval model (RM) as in [Kleinherenbrink et al. 2021], and the low-pass filter (Kf) is applied for comparing results of the initial denoising approach (UM) and the new one (U3): UN = RM(DN); UM = Kf(UN, U); U3 = RM(D3) Figure 1 shows that velocities reconstructed from denoised Doppler shift (U3) seem cleaner than the denoised velocities (UM) and have a lesser effect of thermal noise range variations. Figure 1. Velocity components after the steps of processing: U – initial range and azimuth components, UN – from raw noisy Doppler shift, UM – Kleinherenbrink method, U3 – new smoothing, U4 – smoothing and clustering. The next step of denoising the velocity fields is clustering. The clustering can be described as grouping of objects (in our case pixels) with similar characteristics (in our case velocities and coordinates). The clustering is applied under assumption that sea ice deforms as a solid body with low elasticity and ability for brittle break-up. Therefore, the neighbour elements can have either the same velocity (when they belon to an unbroken ice, i.e., an ice floe) or differ substantially (when they belong to different ice floes). After the clustering is performed, the small-scale variability on the edges of clustering is reduced by applying a median filter to the image with labels (Figure 1). Computing deformations The discontinuities (dUM1) in the field of UM were identified the same way as suggested in [Kleinherenbrink et al. 2021]. This method was slightly optimized for computing discontinuities dUM2 by reducing resolution, smoothing UM with a Gaussian filter and thresholding geometrically averaged Xand Ygradients. Divergence (Ñ) and shear (t) components of sea ice deformation of the velocity fields U, and UA was computed by calculating the velocity gradients (Vg) in neighbour elements and then using equations 23 and 24 from [Kleinherenbrink et al. 2021]. A mosaic of sea ice deformations was created by generating 15 swaths of Harmony observations for the 1 January 2019 and interpolating the swath data onto a grid in Polar Stereographic projection. Daily mean deformation map was created by averaging the individual swaths. Two neXtSIM runs with distinctly different parameters of rheology were used to generate Harmony data, retrieve / denoise velocities and compute deformations. Figure 2 shows that neXtSIM rheology parameters significantly affect the pattern of sea ice deformation. Methods M2 and A were tuned for denoising and computing deformations on the first run and obviously work better with the synthetic data generated from the first run: the major deformation features north of Greenland, Canadian Archipelago and Laptev Sea are well visible. Method A seems to be better capturing also small deformation features and producing a less noisy map. Both methods perform poorer on the data from the second neXtSIM run: large deformation zones north of Greenland, near Novaya Zemlya and in the Laptev Sea are not reconstructed, small features are almost not visible, the maps look quite noisy. Figure 2. Maps of divergence (Ñ) and shear (t) computed from the two neXtSIM runs (upper and lower rows) from original velocities (left) and from Harmony denoised velocities UM2 (middle) and UA (right). Analysis of PDFs of deformation (not shown here) indicates that not only the pattern, but also the PDF of the neXtSIM ice deformation is sensitive to neXtSIM rheology parameters: the second run generally produces lager deformation, and the skewness of PDFs is lower, indicating lesser localisation of deformation zones (wider LKFs). Nevertheless, the PDFs of deformation from two runs (computed either by M2 or by A methods) seem almost equal. That indicates low sensitivity of the noisy ice drift velocity field and corresponding ice deformation to the actual rheology of sea ice and high sensitivity to the processing method. Much of the current understanding of melt ponds (and corresponding sea ice albedo evolution) is based on in-situ studies (e.g. Eicken et al., 1994, Perovich et al., 2012, Tschudi et al., 2008), however, satellite-based observation are the way to map and monitor melt ponds and albedo changes on a pan-Arctic scale. In fact, given that melt ponds are generally small and change rapidly it is of great interest the use of remote sensing data with high spatial and temporal resolution to monitor melt ponds on different types of sea ice during the melting season. The application of satellite data to understand spatial and temporal variations in melt pond coverage has been traditional though estimates using optical sensors (during cloud-free conditions) (e.g. Tschudi et al., 2008, Rösel et al. 2012, Istomina et al., 2020). However, the enhanced spatial resolution of synthetic aperture radar (SAR) compared to radiometers and scatterometers along with SAR’s ability to image our Planet’s surface irrespective of cloud cover, is gaining a lot of interest by researchers to bridge the gap between current knowledge of evolving ice-ocean-atmosphere between melt onset and freeze up (Scharien et al., 2010). However, the availability of spatial resolution to perform large scale monitoring in the high regions is limited. Optical missions like Sentinel-2 provide a very interesting temporal resolution however the availability of cloud-free data is scarce at higher latitudes. Some works have been focusing on the derivation of melt pond fraction and proxy estimates of surface albedo in order to understand the evolution of summer ice albedo and to evaluate the potential of SAR for aiding the parameterization of sea ice and climate models (Scharien et al., 2014). Furthermore, due to the difficulty in collecting data from polar regions, the relatively expensive costs, and logistics, it is important to maximize the potential benefits deriving from data. According to Rolnick et al. (2019), the number of applications of machine learning to study polar regions is not high although it has been increasing over the past decade. The combination of satellite imagery with machine learning holds the potential to address global challenges by remotely estimating environmental conditions in data-poor regions – such as the Arctic region. The proposed approach focuses on using multi sensors to increase the temporal resolution and enhance spatial resolution and accuracy of previous melt ponds and melt pond fraction products or studies. A main focus will be given to synthetic aperture radar (SAR) data in different incidence angles for various ice types. The main reason for using SAR is due to its weather independence, meaning it is not constrained by cloud coverage as optical imagery - which has been the main source of information for melt ponds monitoring. In addition, the heterogeneity of other datasets will be merged with SAR data due to its potential to extent help bridging possible gaps of non-existence data to guarantee the continuous monitoring and understanding of melt ponds formation and dynamics during melting period. The multi source datasets will be used to train models to achieve the following key objectives: 1) Classification of melt ponds 2) Prediction of melt pond fraction Three additional additional and major expected outcomes or by products are: 1) The capability of tracking evolution of melt ponds, as well as extracting the feature importance (i.e. for the case where random forests are applied) to understand the importance of each variable on the prediction results. 2) The creation of training datasets specifically focused on melt ponds, which to our knowledge are not currently available. 3) The retrieval of statistical information such as the relationship between the polarimetric signatures and microwave emissivity of melt ponds, and retrieve information on spectral and radiometric properties, along with other variables that impact melt pond formation. References Tschudi, M. A., Maslanik, J. A. and Perovich, D. K.: Derivation of melt pond coverage on Arctic sea ice using MODIS observations, Remote Sensing of Environment, 112 (5), 2605-2614, https://doi.org/10.1016/j.rse.2007.12.009, 2008. Eicken, H., T. C. Grenfell, D. K., Perovich, J. A. Richter-Menge, and Frey, K.: Hydraulic controls of summer Arctic pack ice albedo, J. Geophys. Res., 109, C08007, https://doi.org/10.1029/2003JC001989, 2004. Perovich, D. K., and Polashenski, C. (2012), Albedo evolution of seasonal Arctic sea ice, Geophys. Res. Lett., 39, L08501, doi:10.1029/2012GL051432. Rösel, A., Kaleschke, L. and Birnbaum, G.: Melt ponds on Arctic sea ice determined from MODIS satellite data using an artificial neural network, The Cryosphere, 6 , pp. 431-446 . doi: http://doi.org/10.5194/tc-6-431-2012, 2012. Istomina, L., Marks, H., Huntemann, M., Heygster, G., and Spreen, G.: Improved cloud detection over sea ice and snow during Arctic summer using MERIS data, Atmos. Meas. Tech., 13, 6459–6472, https://doi.org/10.5194/amt-13-6459-2020 , 2020a. Scharien, R. K., Hochheim, K., Landy, J., and Barber, D.G.: First-year sea ice melt pond fraction estimation from dual polarisation C-band SAR – Part 2: Scaling in situ to Radarsat-2, The Cryosphere, 8, 2163-2176, https://doi.org/10.5194/tc-82163-2014, 2014. Rolnick, D., et al.: Tackling Climate Change with Machine Learning, Association for Computing Machinery, 55, 0360-0300, https//doi.org/10.1145/3485128, 2019. Quadruple Helix Framework for Sea Ice Monitoring: Next Steps Ekaterina Kim, Roger Skjetne, Knut Høyland Norwegian University of Science and Technology, Trondheim, Norway The Arctic - one of the most desolate and sparsely populated areas on our planet - plays a crucial role in regulating the world's climate. Despite Arctic’s importance for the Earth’s climate, this area remains severely under-monitored and not well known to the broad public. Few humans are directly involved in monitoring (this is mostly done using remote sensing tools). As urbanization continues, there is limited human engagement, and participation in monitoring of nature and a further disassociation with it. In view of this challenges, we readopted a quadruplex helix framework for monitoring of sea ice. The framework relates sea ice knowledge at different spatial and temporal resolutions to each other (i.e., remote sensing, in-situ scientific measurements, citizen science, and indigenous knowledge) In addition to the challenges with local data collection, access to (and processing) the Arctic data is a persistent and important multidimensional problem; not all Arctic countries are willing to share their knowledge, but for those who do share, the amount of shared data now outpaces experts’ ability to check, interpret, and make complex decision based on these data. As a result, our ability to predict rapid changes of the environment and manage the effects of environmental change is severely limited. Traditional remote sensing techniques (using satellites) suffer from significant uncertainties near the coast, close to the sea ice edge, etc. Uncrewed aerial vehicles and sail drones have limited coverage and operability window (e.g., darkness, heavy snow, etc.). We are not always able to validate & trust the Arctic satellite products when looking at rapid & localized events or changes on the ground − a recognized shortcoming by the Arctic Council, Copernicus, and International Ice Chart Working group communities. Solving this shortcoming has been difficult in the past due to the lack of technology for automated collection, processing, and quality control of the ground-truth Arctic data. Empirical testing of the popular state-of-the-art machine learning algorithms applied to sea ice revealed a need for pre/post training and systematic evaluations. Fig. 2 is a typical example of false negative errors in some of the analysed algorithms. In addition, sea ice data users (fisheries, forecasting agencies, etc.) increasingly require spatially explicit information on the uncertainty of sea ice parameters for evaluating the risk that a specific outcome of further analysis of the information will be incorrect. However, error estimates are lacking in practically all the sea ice datasets available today. It is, therefore, a priority to develop and standardize methods to compute consistent and comparable error estimates. This is Fig. 1. Quadruplex helix framework for sea ice monitoring and its challenges. particularly important for in-situ observations since realistic uncertainty estimates are essential for meaningful integration of these data in remote sensing and other higher-level products and studies (automated sea ice charting, climate modelling, etc.). Fig. 2 False negative output errors (ChatGPT Jan 9 and Feb 13 versions; Google search engine) During the workshop, the first significant results from two (or more) parallel projects will be presented with the focus on the blue colored elements in Fig. 1. Digital Sea Ice As a part of this project, a multiscale digital method and a system is being built by integrating regional sea ice forecasting models and local ice-ice/ice-structure numerical models with in-situ, shipboard, and space observations of the Arctic sea ice and of environmental conditions. This enables improved spatial and temporal resolution in the models, to achieve more precise forecasting of ice conditions in the Arctic – including better understanding of long-term variations in the polar ice cover. It involves development of novel methods for use of artificial intelligence (AI)-based analytics of synthetic aperture radar and optical imagery from satellites, marine radars and lidars, visual and infrared cameras, and other enabling technologies. Further objectives are to accurately map the sea ice flow in high resolution and generate quality-controlled sea ice drift forecasting. Novel methods for monitoring and analysis of sea ice dynamics and fracturing processes based on data from heterogenous sources are being developed and will be used to update the multiscale model from the real observations. The project aims at developing novel methods and a digital infrastructure for improved spatial and temporal forecasting and decision support in an increasingly dynamic Arctic environment due to climate changes. Such infrastructure will enable more accurate data and information to be produced, thus resulting in better insight on polar Earth systems, as well as improved safety for maritime voyages. NTNU Oceans Pilot on Arctic marine environment This project covers a university-wide aspects of topics from engineering, physics, biology to social anthropology. The Arctic marine environment (including Baltic and Europe’s ice-covered inland waterways) with sea ice and icebergs distinguishes itself from open water environment further south in two vital ways: a) It is more complicated (e.g., more parameters are required to describe it) and b) the available data is scarce. We are working with models and methods to predict short - and long-term behavior of the environment and our human interaction with it, within the framework of risk, reliability, and data. One of the main outcomes is that there are important challenges (and possibilities) in combining measurements in different spatial and temporal scales. Data measured in-situ is usually very local (either in-situ field work with point measurements in space and time, or drifting buoys with continuous temporal records), whereas satellites based usually covers larger areas with a low frequency. In addition, the physics of how the diverse signals from satellites reflect from the ice cover is not completely understood. The uncertainties should be studied through collocated measurements with different sensors (from down on the ice to satellite sensors) and data analysis methods, but also through studying the physics of how signal reflect from different ice covers (wet/dry snow, melt ponds, etc.). During the workshop, we want to engage in dialog with ESA and other members of the program committee and chairs on specific needs, priorities for in-situ observations of sea ice, functional requirements to the collected in-situ data (above, at, and below the sea ice surface) as well as for uncertainty representations, in view of the sensing technologies, data processing and quality control tools being developed at the Norwegian University Science and Technology (NTNU). This will help in achieving a broader impact from ongoing research projects, create the potential for researchers to collaborate across multiple dimensions of scale, and to build a more precise picture of harsh and unforgiving Arctic environment. This is urgently needed in a view of accelerating climate-change-driven events and impacts of human activities on our planet’s ecosystems. Sensor Synergy Sensor Synergy Taillade, Thibault; Engdahl, Marcus; Fernandez, Diego Can We Retrieve Sea Surface Salinity with SAR Measurements? Zhang, Biao; Perrie, William; Zhang, Mingyu Polar Low Recognition and Tracking from Multi-Temporal Synthetic Aperture Radar and Radiometer Observations Rascle, Nicolas Gilles; Grouazel, Antoine; Mouche, Alexis; Nouguier, Frederic; Villarreal Olavarrieta, Carlos Eduardo; Mora Escalante, Rodney Eduardo; Diaz Mendez, Guillermo; Chapron, Bertrand; Ocampo-Torres, Francisco J. Satellite Measurement Of Waves And Currents: SAR Vs Optical Sensors Holt, Benjamin; Lockhart, Brittany; Porter, Mitchell; Comer, Douglas Fronts, Eddies, and Other Features in the Western Pacific and Micronesia from SAR and Other Multi-Sensor Data Husson, Romain; Ollivier, Annabelle; Chehade, Bassam; Peureux, Charles; Quet, Victor; Goimard, Gaël; Soulat, François; Tourain, Cédric; Lachiver, Jean-Michel Comparing and Combining S1 and SWIM Spectral Wave Measurements has a swell peak period around 11 s and an wind-sea peak period around 6 −7 s. The optical spectrum has a wind-sea peak period about 7 −8 s and a swell peak period barely noticeable. The wave directions are also slightly shifted. Reproduction of those effects in numerical simulations (R3S, Nouguier, 2019) is under way within the SARWAVE ESA project (2022-2025). 2.2. Optical imagery to retrieve wave phase speed and current Time series of optical images can help to retrieve wave phase speed and its deviation due to current. Several strategies are investigated, focusing either on sunglint or out of glint images. The wave decorrelation time is investigated using satellite videos of the sea surface. This will provide indications of the ideal viewing geometry of optical sensors dedicated to surface currents, depending on their available horizontal resolution. Confrontation with numerical simulations (R3S, Nouguier, 2019) will be reported within the ESA DopVisSat project (2023-2024). 3. Perspectives This work is part of an effort towards a more systematic use of optical imagery to improve understanding of SAR imagery. Existing optical sensors like Sentinel-2 are considered, and future optical missions with different viewing geometry are also proposed to provide systematic means of calibration and validation of waves and current radar data (e.g the STREAM-O optical companion of STREAM-R mission EE11 proposal). References Chapron, B., Collard, F., Ardhuin, F., 2005. Direct measurements of ocean surface velocity from space: Interpretation and validation. J. Geophys. Res 110 (C07008). Graber, H. C., Terray, E. A., Donelan, M. A., Drennan, W. M., Van Leer, J. C., Peters, D. B., 2000. Asisa new air–sea interaction spar buoy: Design and performance at sea. Journal of Atmospheric and Oceanic Technology 17 (5), 708–720. Janssen, P., Alpers, W., 2006. Why sar wave mode data of ers and envisat are inadequate for giving the probability of occurrence of freak waves. In: Proceedings of SEASAR 2006, SP-613. ESA, ESA - ESRIN, Frascati, Italy. Kudryavtsev, V., Yurovskaya, M., Chapron, B., Collard, F., Donlon, C., 2017. Sun glitter imagery of ocean surface waves. part 1: Directional spectrum retrieval and validation. Journal of Geophysical Research: Oceans 122 (2), 1369–1383. Mari´ e, L., Collard, F., Nouguier, F., Pineau-Guillou, L., Hauser, D., Boy, F., M´ eric, S., Sutherland, P., Peureux, C., Monnier, G., et al., 2020. Measuring ocean total surface current velocity with the kuros and karadoc airborne near-nadir doppler radars: a multi-scale analysis in preparation for the skim mission. Ocean Science 16 (6), 1399–1429. Nouguier, F., 2019. Remote sensing spatial simulator (r3s). Tech. rep., Tech. Rep. SKIM-MPRC-TN6-V1. 0-LOPS-2019, European Space Agency, Noordwijk . 3 Abstract Submitted to ESA SeaSAR2023, Svalbard, May 2-6, 2023 Fronts, Eddies, and Other Features in the Western Pacific and Micronesia from SAR and Other Multi-Sensor Data Benjamin Holt, Brittany Lockhart, Jet Propulsion Laboratory, California Institute of Technology Mitch Porter and Douglas Comer CSRM Foundation Abstract Frontal features, indicating convergence along current shear boundaries and temperature boundaries, have often been observed by SAR imagery as enhanced lines of radar brightness. Conversely, the observation of marine slicks formed by convergence into narrow lines of reduced radar brightness indicates underlying surface flow of small eddies. Both types of convergence features have been observed in boundary currents such as the Gulf Stream, in coastal regions, and associated with islands. A highly cited study by Yoder et al. (1994) made use of astronaut sun glint imagery showing a linear feature in the equatorial Pacific and confirmed that the feature was located along a frontal boundary between warm waters and an upwelling zone. In this presentation, we will discuss frontal features and eddies seen with SAR imagery and multisensory data related to coral-reef islands and open-ocean features in the western Pacific. This study is related to understanding migration patterns of native Pacific Islanders who populated the scattered island groups using traditional sailing techniques, primarily based on a detailed knowledge of navigating by stars and wave patterns. The primary area of interest is the Federated States of Micronesia, centered on the recently-designated World Heritage archaeological site of Nan Madol, located on the island of Pohnpei. To gain understanding of the fundamental ocean conditions and wave patterns as well as climatic forcing events in this region that may have impacted navigation routes and the development of this unique archaeological site, we examined SAR imagery in combination with ocean currents, sea surface temperature ocean color, and wind products derived from satellite data. We identified a unique collection of SAR imagery from JAXA’s ALOS-1 L-band SAR mission, collected in disparate locations in the western Pacific. Current convergence zones of enhanced, bright lines in the open ocean were observed, often away from nearby islands or strong boundary currents, which identify zones of enhanced surface roughness. We also identified dark, curvilinear marine slicks, composed of biogenic films, that are well-known to serve as tracers of the underlying current flow, particularly eddies of various scales. To determine if the frontal features are associated with strong temperature fronts, current gradients and shear, or a combination of both, we overly the SAR features onto coincident SST and current vector maps and then derive backscatter changes of the fronts in relation to surrounding ocean (Figure 1). Returning to examining Micronesia’s volcanic, coral reef island of Pohnpei, a substantial collection of Sentinel-1 imagery was analyzed. These data were obtained in standard beam mode and provided detailed views of waves and refraction patterns, along with fine scale circulation features, including eddies associated with the island. Wind speed measurements were derived using the NOAA Coastwatch SAR winds algorithm (thanks to Chris Jackson). We will discuss steps taken to understand the seasonal patterns of winds, waves, and detailed island SAR features (Figure 2) associate with Pohnpei and an adjacent atoll. Figure 1. a) ALOS-1 SAR image showing enhanced convergence line. b) The linear features are traced and overlaid onto the sea surface temperature map of the same day, showing the alignment of the SAR feature along a filament; c) SAR features overlaid on the same SST map with ocean current vectors indicated. Figure 2. Sentinel-1 SAR image of Pohnpei, Micronesia. The enlargement on the right shows both a bright convergence line likely related to shallow bathymetry as well as small, anticyclonic, bright eddies related to flow through reef openings. Comparing and combining S1 and SWIM spectral wave measurements Husson Romain1, Annabelle Ollivier1, Bassam Chehade1, Charles Peureux1, Victor Quet1, Gaël Goimard1, François Soulat1, Cedric Tourain2, Jean-Michel Lachiver2 1Collecte Localisation Satellites (CLS), Brest, France 2CNES, Toulouse, France Abstract SWIM is a rotating Radar onboard the Chinese-French CFOSAT satellite providing directional wave spectra measurements. This sensor has been operating for 4 years and well complements Sentinel-1 wave measurements due to their different limitations. The loss of Sentinel-1B in December 2022 has reduced the global wave sampling capability with Sentinel-1 constellation and highlighted the relevance of combining these wave measurements. The present study uses the Sentinel-1 Level-2 OCN products from the Wave Mode acquisition mode and the SWIM Level-2P products distributed by ESA and CNES/CLS, respectively. The first part highlights the two sensors complementarities in terms of spatial coverage, sensitivity to varying wave regimes and directional limitations. The second part focuses on their commonly observed wave regimes, for waves with peak wavelength within ~200-500m. The swell measurement performances are assessed using model and in situ measurements and also co-locations between Sentinel-1 and SWIM measurements (also known as cross-overs in altimetry). This involves both static and dynamic co-locations where waves are propagated using a simple linear wave propagation model over a few hundred kilometers to maximize the number of co-locations. These comparisons show significant differences and suggest a need to propose an inter-calibration in order to propose a global and combined swell measurement product. 1. Introduction Ocean surfaces are widely imaged by satellite altimeters providing a dense global coverage with more than 8 altimeters flying operationally today and whose wave products are distributed by the Copernicus Marine Service. These satellites only capture the significant wave height, the integrated energy of the co-existing wave systems. For long, the only sensors able to provide a more detailed sea state description have been Synthetic Aperture Radars (SAR) flying. The SWIM instrument, flying onboard the CFOSAT mission, carries a nadir pointing altimeter on top of its rotating radar and has provided a different point of view on the sea state description [1] with respect to Sentinel-1 with whom it has been operating since its 2018. They both observe the ocean surface and can image the wave spectra at global scale using a synthetic and rotating Radar. Yet, their sensing technology and their global coverage have huge impacts on their capability to observe waves in varying sea state conditions. 2. Data The present study uses the Sentinel-1 Level-2 OCN products acquired in Wave Mode and the SWIM Level-2P products. They are distributed by ESA and CNES/CLS, respectively. They are also compared to directional wave spectra estimation provided by WAM numerical wave model outputs estimated at SWIM measurement's location. Finally, they are also compared to swell partitions extracted from directional wave spectral measurements from in buoys provided by the NDBC network (National Data Buoy Center). Figure 1: Co-located spectra observed with SWIM 10° beam (left), estimated with WAM and observed with Sentinel-1 3. SWIM and S1 specificities and complementarities for observing wave spectra a. Wave classification To simplify the analysis, the wave conditions are split in 3 different sets of waves according to their peak wavekength: [0-200m] corresponding to wind sea dominated sea states (Class 1), [200-500m] to a mix of long wind seas and swells (Class 2) and [500-800m], to very long swell (Class 3). Figure 2: Percentage of occurrence and geographical distribution of the three wave classes according to WAM wave partitions: Class 1 on the left, Class 2 in the middle and Class 3 on the right for the time period April-May-Jun 2021. b. SWIM and SWIM limitations Figure 3: Geographical distribution of Sentinel-1 (left) and SWIM (right) depicted using a 2D histogram of the number of valid partitions acquired at global scale for April-May-Jun 2021. Compared to SWIM, S1 misses some areas mainly near coasts and North Altlantic due to the change in acquisition mode (switching to Interferometric Wide Swath mode - IW). The analysis of the SWIM and S1 data over the various wave classes indicates that for SWIM: - The wave sampling for the longest waves has been limited in the CNES L2 SWIM products to 500m. This threshold was chosen in order to avoid contributions from non-wave phenomena in the ocean waves spectra, possibly. The WAM analysis, indicates that this represents ~10% of the total number of wave partitions. Yet, the longest swell are also the ones associated with most intense storms. They can also greatly affect the sensing capabilities from Doppler altimeters or SWOT (Sea State Bias [3]). - A higher speckly in the along-track direction, which required a very specific processing in order to compensate for the increased noise and energy in this wave spectra region. - The presence of so-called “parasitic peaks”. They correspond to energy peaks becoming relevant whenever looking at wave height spectra while they are relatively weak in SWIM L2P products provided as slope spectra. They are associated with low Signal-to-Noise ratio. In Figure 4, they appear as outliers with a strong difference between the slopeand the heightestimated peak wavelength. For the other partitions, an average bias of 15 m exists. For Sentinel-1, the main limitation lies in the azimuthal cut-off which prevents from imaging waves with azimuthal wavelength shorter than 200m on average. Overall, the SWIM and Sentinel-1 appear best suited for the Classes 1-2 and Classes 2-3, respectivly. The rest of the study focuses on the capability to accurately image wave for this common Class 2 wave category. Figure 4: Scatterplots of peak wavelength estimated from wave height spectra wrt. estimation from slope spectra for a given wave partition domain (left) and histogram of the differences. Only the SWIM most energetic wave partition is considered here. 4. Merging SWIM and S1 swell measurements a. Using S1 classification S1 wave mode acquisition consists in high-resolution 20x20km imagettes that can be classified in metocean phenomena [4]. The estimation of SWIM performances wrt. WAM has demonstrated that co-localized and classified Sentinel-1 imagettes could be used to identify the metocean phenomena degrading SWIM measurements: rain cells, low winds and biological slicks (sea ice not considered). Figure 5: Illustration of the 10 metocean classes of TenGeo-P [4] b. S1 vs. SWIM Using short propagation SWIM and S1 observations (<48h) from their observation location, wave partitions can be compared. This is shown her on Figure 6. A significant underestimation of the SWIM partition peak wavelength is visible wrt. S1 (WV1 and WV2 measurements are mixed). This may be explained by the fact that the S1 azimuth cutoff limitation is not duplicated on SWIM spectra (-24m). This can also be partly explained that the SWIM partition peak wavelength is estimated from the slope spectra while S1 is estimated from the wave height spectra (cf. Figure 4). a. S1 and SWIM vs. Buoy Similarly, comparisons between S1, SWIM and buoys show a consistent difference for the peak wavelength, possibly due to the slopeversus wave height estimation. 5. Conclusions Will be finished later 6. References [1] Hauser, D., C. Tourain, L. Hermozo, D. Alraddawi, L. Aouf, B. Chapron, A. Dalphinet, et al. “New Observations From the SWIM Radar On-Board CFOSAT: Instrument Validation and Ocean Wave Measurement Assessment.” IEEE Transactions on Geoscience and Remote Sensing 59, no. 1 (January 2021): 5–26. https://doi.org/10.1109/TGRS.2020.2994372. [2] Wang, He, Alexis Mouche, Romain Husson, Bertrand Chapron, Jingsong Yang, Jianqiang Liu, and Lin Ren. “Quantifying Uncertainties in the Partitioned Swell Heights Observed From CFOSAT SWIM and Sentinel-1 SAR via Triple Collocation.” IEEE Transactions on Geoscience and Remote Sensing 60 (2022): 1–16. https://doi.org/10.1109/TGRS.2022.3179511. [3] Dubois, Pierre, and Bertr Chapron. “Characterization of the Ocean Waves Signature to Assess the Sea State Bias in Wide-Swath Interferometric Altimetry.” IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, 2018, 3789–92. https://doi.org/10.1109/IGARSS.2018.8518813. [4] Wang, Chen, Alexis Mouche, Pierre Tandeo, Justin Stopa, Nicolas Longépé, Guillaume Erhard, Ralph C. Foster, Douglas Vandemark, and Bertrand Chapron. “A Labelled Ocean SAR Imagery Dataset of Ten Geophysical Phenomena from Sentinel‐1 Wave Mode.” Geoscience Data Journal 6, no. 2 (November 1, 2019): 105–15. https://doi.org/10.1002/gdj3.73. Figure 6: Scatterplot of peak partition wavelength between dynamically co-located SWIM wrt. S1 wave partitions Methodology and Techniques Methodology and Techniques Tings, Björn; Pleskachevsky, Andrey; Wiehle, Stefan; Jacobsen, Sven Ship Wake Detectability in TerraSAR X, CosmoSkymed, Sentinel 1 and RADARSAT 2 Imagery – Summary and Applications for Wake Detection Arthurs, David Open and Reproducible Science: The Role of Computing Platforms for Research and Applications that use SAR Gade, Martin; Peters, Sebastian; Schäfers, Simon On the Use of SAR Data to Monitor Coastal Erosion and Morphodynamics in Intertidal Areas Hajduch, Guillaume; Pinheiro, Muriel; Valentino, Antonio; Vincent, Pauline; Recchia, Andrea; Cotrufo, Alessandro; Franceschi, Niccolo; Piantanida, Ricardo; Benchaabane, Amine; Peureux, Charles; Husson, Romain; Schmidt, Kersten; Mouche, Alexis; Grouazel, Antoine; Nouguier, Frédéric; Johnsen, Harald; Hindberg, Heidi; Guiton, Gilles; Collard, Fabrice Sentinel-1 product performance Shamshiri, Roghayeh; Eide, Egil; Rangriz Rostami, Fazel; Vilhelm Høyland, Knut Sentinel-1 Extra Wide Thermal Noise Removal Using a Deep Learning Model Yitayew, Temesgen Gebrie; Grydeland, Tom; Larsen, Yngvar; Engen, Geir Processing of High Squint Bistatic SAR Data: The Case of Harmony Hoffman, Lauren Alexandra; Mazloff, Matt R; Gille, Sarah T; Giglio, Donata; Bitz, Cecilia M; Heimbach, Patrick Machine Learning for Evaluating the Drivers of Variability in Arctic Sea-ice Motion. Grydeland, Tom; Yitayew, Temesgen Gabrie; Larsen, Yngvar;DuBois, Pierre; Armstrong, Thomas; Gombert, Baptiste; Soulat, Francois; Monnier, Goulven; Hellouvry, Yann-Herve; Camus, Benjamin; Lopez-Dekker, Paco; Lajas, Dulce; Rommen, Bjorn; deWitte, Erik Signal Processing for Harmony: Illustrations with Simulated Data over Ocean Colin, Aurélien; Tandeo, Pierre; Husson, Romain; Fablet, Ronan; Peureux, Charles MediSAR: An Exhaustive Augmented Dataset Of Segmented Sentinel-1 SAR Ocean Observations Of The Mediterranean Sea and the Black Sea regions Stopa, Justin E.; Foster, Ralph; Vandemark, Doug; Wang, Chen; Chapman, Jonathan; Glaser, Yannik; Sadowski, Peter; Mouche, Alexis; Chapron, Bertrand Using Ocean Surface Imagery to Estimate Atmospheric Boundary Layer Stratification Foster, Ralph; Mouche, Alexis; Chapron, Betrand Using SAR Imagery to Diagnose Tropical Cyclone Boundary Layer Mean State Larsen, Yngvar; Engen, Geir; Grydeland, Tom; Yitayew, Temesgen G. An Alternative Approach for Estimation of Doppler Centroid Anomaly Based on Level-0 SAR Data Alpers, Werner R.; Bignami, Franceco Sar Observation Of Internal Waves Generated By Sub-mesoscale Features In The Strait of Sicily Ship Wake Detectability in TerraSAR-X, CosmoSkymed, Sentinel-1 and RADARSAT-2 Imagery – Summary and Applications for Wake Detection Björn Tingsa, Andrey Pleskachevskya, Stefan Wiehlea, Sven Jacobsena a DLR, Maritime Safety and Security Lab Bremen, Am Fallturm 9, 28359 Bremen, Germany [email protected] 1. Introduction Ship wakes are produced by the interaction of the ship’s hull with the ocean water and are result of multiple interacting wave systems closely beneath and on the ocean surface. The ship wake signatures in SAR imagery consists of various components. The most frequently encountered wake components are Kelvin wake arms, V-narrow wake arms and two parts of the turbulent wake: the near field and the far field [1]. The detectability of these four most important wake components in SAR imagery is influenced by several physical variables, which are in the following called influencing parameters. The influencing parameters can be categorized into ship properties, environmental conditions and SAR acquisition settings. In a series of preceding studies of the authors [1, 2], the characteristics of the effects of influencing parameters on the detectability of individual wake components have been modelled using machine learning, categorized, and contrasted against the published state-of-the-art. For the latest study [3], the list of the satellites was extended and the detectability of wake components was investigated in terms of different radar frequency bands (C-Band and X-Band SAR) and different orbit altitudes (i.e. slant ranges). This study summarizes the method and the results of the preceding studies [1,2,3] and the application of the results to the actual task of wake detection is demonstrated. The demonstration shows that the developed models can be applied to control the precision performance of wake detectors and to estimate vessel velocity with an accuracy coinciding with other published methods [4]. 2. Materials and Method The studies are based on four different SAR missions (Table 1). The ground truth wake samples listed for each sensor were created by a manual inspection procedure. Table 1: Summary of wake component datasets Sensor name TerraSAR-X (TSX) CosmoSkymed (CSK) Sentinel-1 (S1) RADARSAT-2 (RS2) Frequency band /radar wavelength [cm] X / 3.1 X / 3.1 C / 5.6 C / 5.6 Orbit-Altitude [km] 514 619 693 798 Approx. slant range [km] at 30°/50° incidence angle 593 / 800 715 / 963 800 / 1078 922 / 1242 Acquisition modes / product types SL, SM / MGD HIMAGE / DGM IW / GRDH MF, F, S / SGF Number of total wake samples (HH / VV) 2881 (2429 / 452) 94 (94 / 0) 618 (0 / 618) 407 (407 / 0) The detectability of each of the four wake components was modelled for each of the four sensors using the support vector regression (SVR) method. The length of each wake component is used as indicator for the wake component’s detectability. The preceding studies have an intersecting set of five influencing parameters, which are listed in Table 2. The SVR models predict which wake component lengths are expected depending to the conditions defined by the influencing parameters. The predicted wake component lengths are then linearly normalized between a minimum and maximum length boundary to obtain a measure of detectability with uniform scale. The socalled detectable length metric (DLM) for a sensor 𝑠 and a wake component 𝑤 is: 𝐷𝐿𝑀𝑤,𝑠(𝑥1,… ,𝑥5)=(𝑓 𝑤,𝑠(𝑥1,…,𝑥5)− 𝑙𝑤 𝑚𝑖𝑛) |𝑙𝑤 𝑚𝑎𝑥 − 𝑙𝑤 𝑚𝑖𝑛| ⁄ (Eq. 1) where 𝑥1,…,𝑥5 defines the five influencing parameters and 𝑓 𝑤,𝑠 the SVR model. 𝑙𝑤 𝑚𝑖𝑛 is the minimum length boundary and 𝑙𝑤 𝑚𝑎𝑥 maximum length boundary, both depending on the respective wake component. Harmony end-to-end performance simulator: evaluating the performance of a bi-static ATI SAR mission for ocean observations. P. Dubois1, G. Monnier2, P. Lopez-Dekker3, T. G. Yitayew4,T. Armstrong1, B. Gombert1, F. Soulat1, Y-H Hellouvry2, B. Camus2, T. Grydeland4, D. Lajas5, B. Rommen5, E. deWitte5 (1CLS, 2Scalian DS, 3TU Delft, 4NORUT, 5ESA-ESTEC) On September 22, 2022, the European Space Agency announced Harmony as the 10th Earth Explorer (EE10) mission. The Harmony concept comprises of two identical satellites that will fly in a constellation with a Copernicus Sentinel-1 satellite ([1]). Each satellite is being designed to carry a receive-only synthetic aperture radar as its main instrument. Working together with Sentinel-1’s radar, Harmony will provide simultaneous measurements of wind, waves, and currents. These, along with measurements of sea surface thermal difference and cloud motion, will enable an unprecedented view of the marine atmospheric boundary layer. The ocean products performance, derived in the early development phase with models and targeted simulations, was consolidated through the use of end-to-end (E2E) simulations, following ESA’s proven process for EE selection ([2]). The E2E simulators are a classic tool for characterizing the performance of a mission, as defined by the science requirements. They integrate the definition of a set of geophysical truths, the geometry and timing of the acquisition, the transfer function of the instrument, and the prototyping of all levels of processing (On-board L0, L1, L2). At the end of the L2 processing, the estimates of the geophysical parameters of interest can be compared to the geophysical truth sets used as an input to the simulation. The radar integration, modeled by the radar equation, is supported by a numerically processed spatial integral, whose complexity of computation is related to the spatial and temporal sampling of the surface, to the representativeness of radiometry and instrument model (antenna pattern, . . .). In the context of a SAR system allowing, among other things, the estimation of surface currents vectors (with the ATI capability), the simulation of raw data requires a fine description of the ocean surface both spatially and temporally. Adding the number of phase centers, the variety of received polarization, and the extent of the scene (especially when Sentinel-1 emits in the IW mode), the resulting computational time is considerable and is the main challenge of the E2E design (called HEEPS/Mare for ocean application). The choice of models and implementation strategies made it possible to achieve a data production time compatible with performance studies. 1 Ocean scene modeling The ocean scene is considered as a collection of layers, each consisting of a mesh that is defined at given spatial resolution and animated, if needed, at given temporal resolution. The layers are: for the atmospheric conditions, the (spatially variant) (a) wind conditions, (b) clouds, wet troposphere. for the time varying waves elevations, the (spatially variant) (a) aforementioned wind conditions layer, setting local wind sea elevation spectra by virtue of a modelled relationship between wind speed and friction velocity, (b) swell conditions, setting local swell elevation spectra, (c) surface currents, contributing to set the local waves dispersion relationship. The local waves spectra allow for the generation of local wave realizations thanks to the Fast-FT. Unfortunately, composing the entire surface using the tiles from the FFTs would result in strong discontinuities at the tile boundaries. An innovative approach is used to create smooth transitions between the local realizations. for the surface radiometric model, a set-up that ensures a space-time cross-correlation of the backscattered signal that coincides with a chosen Normalized Radar Cross Section (NRCS) model ([3]) at t= 0 and whose temporal spectrum coincides with a chosen Doppler spectra model ([4]). A mesh approach is applied, with the use of resolved waves elevation in the phase during the radar equation spatial integration, and the remaining contribution from the waves smaller than the mesh resolution (short waves) are accounted using time varying space random complex draws. For each of the mesh facets, the spectrum of the cross-correlation of the random variable is guaranteed to be the Doppler spectrum model computed for short waves only, and similarly for the NRCS. In doing so, it is ensured that the effect of short wave 1 motion on phase and amplitude (including its decorrelation) is taken into account. The setting of this statistical model is put in a table which is function of meshed slopes (that can be computed at the same time than meshed elevations, see previous item), facet size, wind speed and direction. 2 Geometry, Timing and Instrument transfer function The spatially and temporally sampled data are acquired with respect to a specific line of sight pointing towards Earth. For Sentinel-1, the orbit and pointing information is determined using the Earth Observation CFI software and input state vectors. The Harmony satellites are propagated along the same orbit, taking into account a desired separation and the reception time given the bi-static observation geometry. The mechanical pointing is determined to be the same as that of the Sentinel-1 boresight, and the reception is configured based on a given Sentinel-1 acquisition mode. For the reception of the signal from the simulated ocean scene, the complex impulse response of the surface is computed utilizing input antenna models for both emission and reception, with narrow range gates to ensure precise estimate of the returned power. This result then undergoes a circular convolution with the chirp form to provide results expressed in sampling frequency domain, to which thermal noise is added. 3 Processing The processing starts with the focusing of the raw data, as detailed in [5]. The following step is the retrieval of the main ocean products of interest in the relation to the ocean-atmosphere interactions [6]. The Doppler chain. The [L1b] Geophysical Doppler Centroid is estimated by removing the Along-Track Interferometry lag and the geometric phase contribution to the data. For Sentinel-1, this is the Doppler Centroid Anomaly (DCA) phase, as it delivers single phase-center data only, and the ATI phase for Harmony, exploiting the baseline between the wing antennas. The [L2] Relative Total Surface Current is the [L1b] geophysical Doppler where the so-called wave-Doppler contribution has been estimated by Geophysical Model Functions (GMFs) interpolation, and removed. The amplitude chain. The Normalised Radar Cross-Section (NRCS) is, after appropriate calibration, a direct radar observable. The computation of the coand cross-spectra includes the rotation of the polarisation basis, the tiling of the data, and processing steps similar to the S1-OSW processing. The surface stress and stress equivalent surface wind uses a GMFs that relates the stress equivalent surface wind and the NRCS. There are at least three observations at different Lines of sight for the L2 estimates, which makes estimation of vectors possible (and over-determined). Parts of the Harmony L2 processing rely on the use of GMFs to correct/estimate geophysical contributions to the signals (amplitude and Doppler). These GMFs are one of the key elements of the mission performance study. Using GMFs based on real radar data in the E2E context can lead to uncertainties in the conclusions, since the ocean scene modelling is not the real ocean. A safe option has been to construct GMFs using the same models as those used in the E2E simulator. 4 Performance results The main purpose of the simulator is to provide realistic estimates of expected Harmony performance. It can, for example, demonstrate the capacity of Harmony to retrieve a heterogeneous 2D wind velocity field (Figure 1) or a 2D wave spectra (Figure 2). In addition, the simulator allows to quantify performance metrics of L1 and L2 products such as the Noise Equivalent Sigma Zero and the standard deviation in retrieved wind and current homogeneous velocity fields. Initial results have been obtained during Phase A, and will be consolidated with updated processing algorithm definitions and more ocean scene variety. References [1] Harmony, eoportal. https://www.eoportal.org/satellite-missions/harmony. Accessed: 2023-0123. 2 Figure 1: Capability of Harmony to retrieve a complex wind field in WV1 mode simulated using the HEEPS/Mare software. Left: input heterogenous wind field modelled as U10(x, y) = (5 + 0.25x−0.25y)x+ (5+0.5y)y. Right: retrieved wind field with a 125 m x 125 m sampling. Black arrows indicate wind direction Figure 2: Capability of Harmony to retrieve an ocean swell spectrum in WV1 model using the HEEPS simulator. Left: input scene wave height spectrum. Middle and Right: retrieved modulation cross-spectrum for Sentinel-1 and Harmony-A, respectively. The difference in spectral power is explained by the Geophysical Model Function (GMF) that must be applied to the modulation spectra to obtain the wave height spectra. [2] ESA. Report for Mission Selection: Earth Explorer 10 Candidate Mission Harmony. Technical Report ESA-EOPSM-HARM-RP-4129, European Space Agency, Noordwijk, The Netherlands, 2022. [3] A Voronovich. Small-slope approximation for electromagnetic wave scattering at a rough interface of two dielectric half-spaces. Waves in Random Media, 4(3):337–367, 1994. [4] F Nouguier, C-A Guerin, and G Soriano. Analytical Techniques for the Doppler Signature of Sea Surfaces in the Microwave Regime—I: Linear Surfaces. IEEE Trans Geosci Remote Sens, 49(12):4856–4864, 2011. [5] Y. Larsen, Grydeland, T. G. T., Yitayew, and G. Engen. Processing of high squint bistatic sar data: The case of harmony. SeaSAR, 2023. [6] P. Lopez Dekker, M. Kleinherenbrink, A. Payez, A. Stoffelen, and B. Chapron. Harmony Algorithm Theoretical Baseline Document: Oceans and Air-Sea Interactions. Technical report, TU Delft, 2022. 3 WaddenSAR campaign: first results Paco Lopez-Dekker1, Marcel Kleinherenbrink1, Andreas Theodosiou1, Marieke Eleveld2, Firmijn Zijl2, Karlus Macedo3, Thiago Ruiz3, and Julia Kubanek4 1TU Delft, 2Deltares, 3Metasensing, 7ESA February 13, 2023 1 Introduction In September 2022, the Harmony mission [1, 2] was officially confirmed as the 10th ESA Earth Explorer. Harmony’s space segment will consist of a pair of receive only-radar satellites that will fly in formation with Sentinel-1 which will act both as a common radar transmitter and, from a Harmony-mission perspective, as a third radar receiver. During most of the mission one of the Harmony satellites will fly 350 km ahead of Sentinel-1, with the other satellite trailing Sentinel-1 at the same distance. The resulting system will provide simultaneous measurements of the Normalized Radar Cross Section (NRCS) from three directions, implementing roughly the equivalent to a very high resolution scatterometer. In addition, Harmony will provide Doppler measurements using short-baseline Along-Track Interferometry (ATI) in the case of the companion satellites, and falling back to the much less sensitive Doppler Centroid Anomaly technique in the case of Sentinel-1. The large inter-satellite separation, which is required to provide the directional diversity required to retrieve surface stress and surface current vectors, results in unprecedented large bistatic angles. While TanDEM-X [3] has been, from a system perspective, the first multistatic SAR mission, it is safe to say that Harmony will be the first multistatic mission from an electromagnetic scattering point of view. The characteristics of bistatic scattering at the ocean surface have been theoretically studied in depth [4, 5]. However, experimental data is required to fully validate and fine tune these scattering models. To serve that purpose, given the lack of available bistatic observations, ESA organized a multistatic airborne campaign in the Dutch Wadden Sea, with Metasensing as the prime contractor. The main goals of the WaddenSAR campaign were: 1. To demonstrate the Harmony mission concept. 2. To test the proposed airborne implementation for use in future campaigns. 3. To confirm that the assumption that bistatic measurements behave largely like monostatic ones with a monostatic-equivalent geometry. Confirming this assumption allows using monostatic airborne data, which are much easier to collect, as a proxy for bistatic data in the development of, for example, retrieval algorithms. The WaddenSAR flights took place on the 14th and 16th of March, 2022. This paper provides an overview of the campaign set-up and discusses some preliminary results. 2 Campaign description During the campaign, Metasensing flew two Cessna C208 aircraft in a formation emulating Harmony’s multistatic viewing geometry. As shown in Figure 1, the trailing aircraft carried a FMCW C-band radar, transmitting in V-polarization and receiving in two polarizations. The leading aircraft carried a receive-only radar payload with the dual-polarized antennas pitched 29.5◦in order to align the receive-antenna footprint with that of the transmitter. Along-track interferometric observations where achieved by toggling in transmit between a pair of antennas separated about 0.8 meter. The raw data were focused on a regular ground-projected grid aligned with he flight direction using a back-projection algorithm. Table 1 provides a list of some relevant geometric and radar parameters. On the first flight day, about 400 km of data was acquired flying a circuit in an open-sea region, North of the Wadden Islands. On the second day, after solving some technical issues, over 200 km were acquired mostly inside the Wadden Sea. Figure 2 shows the planned acquisition tracks overlayed on the total current predicted by the Dutch Continental Shelf Model in Flexible Mesh (DCSM-FM) model developed and implemented by Deltares. 1 Figure 1: Left: bistatic receive-only radar installed on the leading aircraft, with the dual-pol antenna tilted to provide overlapping antenna footprints. Right: transmit-receive system installed on trailing aircraft, with two transmit antennas toggled in order to implement ATI. 3 Results Very preliminary results (not included in this abstract) show that both the monostatic and bistatic amplitudes are of good quality and generally similar to each other, as expected. The quality of the monostatic ATI phase should be enough to capture some of the strong tidal features shown in the model. At time of writing, the bistatic ATI phase shows large systematic range dependent phases, which the authors hope to have understood and mostly cleaned up during the workshop. Figure 2: Planned flight tracks in the Wadden Sea overlayed on tidally dominated local currents predicted using the Dutch Continental Shelf Model in Flexible Mesh (DCSM-FM) model developed at Deltares. References [1] P. L´opez-Dekker, H. Rott, P. Prats-Iraola, B. Chapron, K. Scipal, and E. D. Witte, “Harmony: an Earth Explorer 10 Mission Candidate to Observe Land, Ice, and Ocean Surface Dynamics,” in IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium, pp. 8381–8384, July 2019. ISSN: 2153-6996. Flight & Observation geometry Flight altitude ∼1550 m Aircraft along-track distance 1100 m to 1400 m Flight speed ∼60 m s−1 Angle of incidence (monostatic) 25◦to 45◦ Swath width ∼1200 m Radar parameters Center frequency 5.3 GHz Azimuth single-look resolution 0.5 m Slant-range single-look resolution 2.8 m ATI baseline 0.8 m Transmit polarization V Receive polarization Dual (H and V) Table 1: Radar and observation geometry characteristics [2] P. L´opez-Dekker, J. Biggs, B. Chapron, A. Hooper, A. K¨a¨ab, S. Masina, J. Mouginot, B. B. Nardelli, C. Pasquero, P. PratsIraola, P. Rampal, J. Stroeve, and B. Rommen, “The Harmony Mission: End of Phase-0 Science Overview,” in 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS, pp. 7752–7755, July 2021. ISSN: 2153-7003. [3] G. Krieger, A. Moreira, H. Fiedler, I. Hajnsek, M. Werner, M. Younis, and M. Zink, “TanDEM-X: A Satellite Formation for High-Resolution SAR Interferometry,” IEEE Transactions on Geoscience and Remote Sensing, vol. 45, pp. 3317–3341, Nov. 2007. [4] T. Elfouhaily, D. R. Thompson, D. Vandemark, and B. Chapron, “A new bistatic model for electromagnetic scattering from perfectly conducting random surfaces,” Waves in Random Media, vol. 9, pp. 281–294, July 1999. [5] L. Iannini, D. Comite, N. Pierdicca, and P. Lopez-Dekker, “Rough-Surface Polarimetry in Companion SAR Missions,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–15, 2022. Proceedings Papers List of Proceedings Papers Korosov, Anton, Malin Johansson, Anja Frost, Juha Karvonen, Ekaterina Kim, Ron Kwok, Nicolas Longépé, Johannes Lohse, Robert Shuchman, Catherine Taelman, and Stefan Wiehle, Recent advances in SAR remote sensing of sea ice and recommendations for the future: Summary from SEASAR 2023. Johannessen, Johnny A., Bertrand Chapron, Fabrice Collard, Ben Holt, Jose da Silva, Artem Moiseev, Werner Alpers, Thibault Taillade, Lucile Gaultier, Biao Zhang, William Perrie, Antonio Bonaduce, Nicolas Rascle, Lasse Pettersson, and Craig Donlon, Satellite Sensor Synergy to strengthen spatial and temporal coverage and shortcut interpretation challenges Stopa, Justin, Ralph Foster, Aurélien Colin, Martin Gade, Lauren Alexandra Hoffman, Roghayeh Shamshir, and Björn Ting, Advancements in methodologies using machine learning – SEASAR2023. Jones, Cathleen E., Maria Michela Corvino, Benjamin Holt, Elena Morando, Roberto Del Prete, Martin Gade, Scott Kaczor, Peter Lanz, Bou-Laouz Moujahid, and Yi-Jie Yang. SEASAR Applications: Status and Outlook. 1 RECENT ADVANCES IN SAR REMOTE SENSING OF SEA ICE AND RECOMMENDATIONS FOR THE FUTURE: SUMMARY FROM SEASAR 2023 Anton Korosov1, Malin Johansson2, Anja Frost3, Juha Karvonen4, Ekaterina Kim5, Ron Kwok6, Nicolas Longépé7, Johannes Lohse2, Robert Shuchman8, Catherine Taelman2, Stefan Wiehle3 (1) Nansen Environmental and Remote Sensing Center, Bergen, Norway (2) UiT The Arctic University of Norway, Tromsø, Norway (3) DLR (German Aerospace Center), Germany (4) Finnish Meteorological Institute, Helsinki, Finland (5) Norwegian University of Science and Technology (NTNU), Trondheim, Norway (6) University of Washington Applied Physics Laborator, Seattle, WA, USA (7) European Space Agency-ESRIN, Rome, Italy (8) Michigan Technological University, USA 1. INTRODUCTION Sea ice plays a crucial role in the Earth's climate system and global ecology. Its presence significantly influences the planet's energy balance by reflecting sunlight, helping regulate temperatures, and influencing ocean circulation patterns. Sea ice also serves as a habitat and a crucial feeding ground for various marine organisms, supporting complex ecosystems from microscopic algae to large mammals such as polar bears and seals. Additionally, the freezing and melting processes of sea ice impact ocean dynamics and, in turn, affect weather patterns and the livelihoods of coastal communities. Importantly, sea ice can be viewed also either as an obstacle for navigation or as an extension of land used by local communities for travel and hunting. Understanding the dynamics and importance of sea ice is essential for predicting climate changes and their potential impacts on both local and global scales. Synthetic Aperture Radar (SAR) remote sensing is a key tool for monitoring and studying sea ice due to its unique capabilities in all-weather and day-and-night imaging. The importance of SAR remote sensing in the context of sea ice lies in its ability to provide high-resolution data, offering detailed information about sea ice distribution, type, and dynamics over vast and often remote regions. SAR sensors can penetrate through clouds and darkness, allowing continuous observation and monitoring of sea ice cover regardless of weather conditions. These observations are fundamental for understanding the changing patterns of sea ice extent, thickness, and movement, enabling the assessment of ice dynamics, identification of different ice types, such as ridged, deformed, or open water areas, and the tracking of ice movement and deformation. Such data is invaluable for climate studies, weather forecasting, fishing and hunting, safe navigation, ecosystem monitoring, and aiding in the understanding of the broader implications of climate change and human activities on polar regions. SAR remote sensing, therefore, plays a pivotal role in providing essential data for scientific research and policymaking concerning sea ice and its dynamic responses to changing environmental conditions. In this work we summarize recent advancements in SAR remote sensing of sea ice presented at the SeaSAR’23 conference and provide a synopsis of the main challenges and recommendations. 2. OVERVIEW OF RECENT ACHIEVEMENTS 2.1 Copernicus Marine Service SITAC SAR-Based Baltic Sea Ice Products The Baltic Sea ice drift product in the Copernicus SITAC is provided by Finnish Meteorological Institute (FMI). The algorithm is a two resolution (referred as low and high resolution) level model. Co-registered dual-polarized (HH/HV) C-band SAR images from Sentinel-1, Radarsat-2 or 2 Radarsat Constellation Mission are first converted to single channel images by computing the SAR backscatter magnitude M=(HH2 + HV2)1/2 from the two calibrated SAR channels. After this ORB (Oriented FAST and Rotated BRIEF, Rublee et al., 2011) is applied to detect ice drift in a reduced resolution of 500 m. This low-resolution result is then interpolated to cover the whole sea ice area. Optical flow (Horn and Schunck, 1981) is applied to high-resolution magnitude images using the low-resolution drift as an initial shift between the two images. The Lucas-Kanade Optical flow (Lucas and Kanade, 1981) is applied in the algorithm. Lucas-Kanade algorithm assumes that the optical flow equation holds for a block of pixels and the motion is solved by a least-squares fit. A simplified algorithm flow diagram is presented in Fig. 1. Figure 1. Structure of the operational FMI SITAC ice drift algorithm. A pair of overlapping images in input and first co-registered and resampled to 100 m resolution, then magnitude images of the two SAR channels are created, they are resampled to 500 m resolution and fed to the ORB feature detection and matching algorithm. The 500 m (low-resolution) drift detection is then interpolated to cover the image sea ice area. The low-resolution SID is then used as a starting point to the optical flow drift detection applied to the full-resolution magnitude images to refine the low-resolution ice drift estimation. 2.2 Tracking backscatter signatures of individual sea ice floes using in-situ drift observations At present, most of the SAR data that is routinely available for sea ice monitoring is acquired in wide-swath mode at C-band, for example by Sentinel-1 (S1). For these sensors, the backscatter signal from a given sea ice type varies with incidence angle (IA) across the swath (Lohse et al., 2020), and it is strongly affected by changes in temperature and snow properties (Barber et al., 1992). Understanding these variations in backscatter is crucial for the interpretation and automated analysis of the imagery. The seasonal evolution of C-band radar backscatter has been extensively studied for landfast ice (for example in (Yackel et al., 2000)), but seems rare for drifting ice. Tracking drifting ice floes in consecutive SAR images over long timespans has proven to work well in winter but remains challenging in melting conditions or in regions characterized by high drift speeds (Krumpen et al., 2019). In this study, we use a set of in-situ drift trajectories collected during the CIRFA-22 campaign to track individual ice floes in the Fram Strait over several months, covering the transition from freezing conditions to warmer melting conditions. We automatically identify the drifter locations in overlapping S1 imagery and manually identify and track distinct surface features (hereafter called ‘ROIs’) in the vicinity of the drifters that can be followed over time in the SAR images. On short timescales (days), this allows us to investigate the IA dependence of the backscatter signature for the exact same sea ice, while on longer timescales (weeks), we can study the temporal backscatter evolution of drifting ice as it undergoes physical changes at melt onset. We investigate the IA dependence by estimating the slope relating IA to backscatter based on the backscatter values of a ROI imaged in two consecutive SAR scenes, once in near-range and 9 L-band SAR was found to be a useful indicator of early melt stages, due to the shift from stronger VV to stronger HH data. The same trend could not be observed in the C-band images, and as such the two frequencies complemented one another. SAR images from the CIRFA cruise in 2022 at the Belgica Bank fast ice was found to confirm this analysis and the usefulness of PD will be further investigated using both the MOSAiC, N-ICE2015 and the CIRFA cruise data, utilizing overlapping Land C-band images that co-inside with in-situ measurements. 2.8 Quadruple Helix Framework for Sea Ice Monitoring: Next Steps Ice covered Arctic areas remain severely undermonitored and not well known to the broad public. Current ecosystem mapping proposals via the EU’s Biodiversity Strategy for 2030 do not explicitly include in-situ measurements as a part of the mapping. Without in-situ monitoring, rapid localized changes in ecosystem's conditions cannot be detected. As a result, our ability to predict rapid changes of the environment and manage the effects of environmental change will remain severely limited. Furthermore, as waves of urbanization and immigration as well as the widespread of digitalization continue, there is limited human engagement, and participation in monitoring of nature and a further disassociation with it. We illustrate the above challenges using a quadruplex helix framework for monitoring of sea ice as an example. The framework (shown in Fig. 7) relates sea ice knowledge at different spatial and temporal resolutions to each other (i.e., remote sensing, in-situ scientific measurements, citizen science, and indigenous knowledge) Users from the maritime sector are not always able to validate and trust satellite products when looking at rapid and localized events or changes in sea ice conditions. This shortcoming can be addressed using technology for automated collection, processing, and quality control of the ground-truth sea ice data. A three-stage approach for the automated analysis of close-range optical images containing floating ice is described in Panchi et al. (2021). The proposed system is based on an ensemble of deep learning models and conditional random field postprocessing. Figure 7. Quadruplex helix framework for sea ice monitoring and its challenges. 10 The following surface ice formations are considered: icebergs, deformed ice, level ice, broken ice, ice floes, floe bergs, floe bits, pancake ice, and brash ice as well as additional five nonsurface ice categories: sky, open water, shore, underwater ice, and melt ponds. The best performance is achieved using an ensemble of models having pyramid pooling layers (PSPNet, PSPDenseNet, DeepLabV3+, and UPerNet) and convolutional random field postprocessing, and this outperformed the best single-model approach. The results show that when per-class performance was considered, the sky is the easiest class to predict, followed by deformed ice and open water. Melt pond is the most challenging class to predict. More efforts will be needed to make this automated image segmentation and analysis fully operational for remote sensing applications—especially the collection and labelling of more images containing floeberg, floebit, and melt ponds and introducing new labels. In situ verification, validation, and possible corrections to ice segmentation would go a long way in providing more training data and improving the proposed approach. When coupled with optical sensors and GNSS, such an approach can serve as a supplementary source of large-scale ‘ground truth’ data for validation of satellite-based sea-ice products. In addition, sea ice data users (e.g., maritime sector) increasingly require spatially explicit information on the uncertainty of sea ice parameters for evaluating the risk that a specific outcome of further analysis of the information will be incorrect. It is, therefore, a priority to develop and standardize methods to compute consistent and comparable error estimates for sea ice datasets. This is particularly important for insitu observations since realistic uncertainty estimates are essential for meaningful integration of these data in remote sensing and other higherlevel products and studies (automated sea ice charting, climate modelling, etc.). 3. REMAINING KNOWLEDGE GAPS AND CHALLENGES In the field of SAR remote sensing of sea ice, several knowledge gaps persist, necessitating further investigation and development. One critical issue is the presence of high thermal or speckle noise. Effective noise suppression techniques that maintain image resolution are essential. The implications of noise on the accuracy of ice drift and deformation measurements need thorough examination. The loss of the Sentinel-1B satellite has reduced data coverage, emphasizing the need for access to RCM data and algorithm adaptation. Thermal noise reduction is particularly needed for the RCM cross-polarization channel due to the strong scalloping effect and significant noise patterns, which affect automated SAR interpretation algorithms. Another challenge is the low contrast in SAR images, particularly in wet snow conditions during summer and broken ice within the MIZ. Enhancing the informativeness of SAR image patches is crucial for robust sea ice drift retrieval based on Maximum Cross-Correlation (MCC) methods. Utilizing HH and HV polarization channels simultaneously may offer a solution, but efficient methodologies need to be developed. Rapidly changing surface conditions, such as melting in summer and heterogeneous drift in the MIZ, complicate image analysis. Adapting the time delta between images is necessary, and identifying the optimal time interval is critical. The inherently different patterns observed in Cband and L-band SAR imagery further complicate MCC and alignment evaluation, necessitating methods to compare and align these datasets effectively. At the same time image alignment algorithms face significant challenges due to floe rotation and other fast surface changes, which complicate image morphing. More efficient image generation and morphing techniques are needed, as well as new metrics for evaluating aligned multi-frequency imagery. Sea ice numerical models need a proper Lagrangian Sea Ice Drift product for thorough 11 calibration and validation. However, despite current SAR missions, such as S1, Radarsat-2 (RS2), and the RCM provide vast amounts of SAR imagery, there is currently no replacement for the ice drift product produced by the Radarsat Ground Processing System (RGPS) (Kwok et al., 1998). Obviously, the handling of large SAR datasets (Big Data) demands optimized algorithms and adequate hardware resources. Sea ice is highly diverse, and some categories, such as new ice, are underrepresented. Improving the representation of all ice types in SAR data and addressing low prediction accuracy in underrepresented categories is critical. Strong winds and the presence of new or young ice further complicate classification, necessitating improved quality of label data and mitigation of systematic biases in ice charts. The role of snow in influencing radar signatures, including snow metamorphism and the effects under dry freezing conditions, must be better understood. Investigating the impact of windcompacted layers, rain-on-snow events, ice lenses within the snowpack, and the brine layer at the snow-ice interface, particularly in relation to Cand X-band SAR, will improve the interpretation of radar data. The L-band's relative insensitivity to these factors should also be considered. Inconsistent classification results across multitemporal images presents another challenge. Developing methods to improve consistency and address these inconsistencies is necessary. Furthermore, machine learning (ML) and deep learning (DL) models, often considered black boxes, require efforts to elucidate the physical relationships between input features and the theoretical informativeness of SAR images to infer multiple ice types. Multi-sensor synergy offers enhanced temporal and spatial coverage when combining data from multiple sensors. However, the time delays between acquisitions, particularly in regions with high sea ice drift speeds like Fram Strait, pose significant challenges. Data alignment is crucial, but temporal gaps in multi-sensor data can be problematic. Combining SAR with optical satellite data may offer advantages for various sea ice tasks, for instance, infrared sensors can differentiate between thin and thick ice by heat fluxes, optical sensors can identify open water, snow-covered sea ice, and ridges under favorable illumination, and SAR can penetrate snow and reveal ice structures, although separating ice from water remains challenging. However, overcoming the time separation between image acquisitions remains a key issue. Upscaling and downscaling between different observational modes, from in-situ to drones, airborne, and satellite data, to models, requires careful consideration of ice motion. Large spatial coverage over study sites can help mitigate issues with overlapping drifting in-situ campaigns, enhancing the overall understanding and monitoring of sea ice dynamics. 4. OUTLOOK AND RECOMMENDATIONS 4.1 Algorithm Improvement To enhance the retrieval of sea ice drift from SAR imagery, the implementation of deep learning techniques is recommended. Post-processing methods such as discarding, optimization, and interpolation of drift vectors should be developed to refine the data quality. Furthermore, the optimization of time intervals between SAR image acquisitions is crucial for improving the accuracy of ice drift measurements. Algorithms must be optimized for parallel processing, and more resources should be allocated for processing. 4.2 More Input Data An operational L-band SAR constellation should be launched to improve data coverage and quality. Collocating datasets from various missions, including passive microwave, altimetry, scatterometry, and SAR, will enable synergetic use and enhance the robustness of sea ice monitoring. Incorporating additional variables into automatic algorithms will improve their accuracy. These variables include wind speed, solar radiation, optical data VIIRS, PMW data from AMSR2, RS2, RCM, high-resolution Sea Surface Temperature (SST), radar and laser 12 altimetry, multi-frequency and multi-polarization data, and sea ice deformation data. Improving time separation between different SAR sensors and leveraging tandem missions like S1 and ROSE-L for automated ice type classification are recommended. A fleet of mixed micro-satellites, such as those from Capella Space, could be an option when time delays of less than one hour are acceptable. The use of RCM mode HH+VV over polar regions in summer, and combining sea ice deformation with thermodynamics for classification, should be explored. Identifying the contributions of L-band SAR for improved sea ice products and enhancing collaboration between different sensor acquisitions will maximize the potential of multisensor synergy. 4.3 Better Training and Validation Data High-resolution verification data from other satellites or in-situ measurements coinciding with SAR data are crucial for accurate training and validation. Ridges, leads, and roughness data from altimetry, which are free from inherent biases in manual ice charts, should be utilized. Building a solid training dataset, including challenging conditions such as wet ice in summer and windy water, with many scenes similar to AI4Arctic, will improve model robustness. Cross-calibration of ice experts through evaluation of ice charts is also necessary. 4.4 Better output Products Increasing the temporal resolution of ice drift products is essential. Developing a highresolution, long-term Lagrangian sea ice drift dataset will provide valuable insights into sea ice dynamics. Additionally, combining sea ice drift data with thermodynamic models can facilitate the retrieval of sea ice thickness, thereby improving the overall understanding of sea ice properties. Satellite data should also be used to derive new variables, such as the probability of belonging to an ice category, deformation, ridges, leads, and aerodynamic roughness. Developing fit-for-purpose ice products tailored for ice charts or for models is essential for providing relevant and actionable information. 4.5 Improved Applications A move towards integrated systems combining satellite observations, data assimilation, and modeling is recommended. The use of Structural Similarity (SSIM) as a metric for evaluating the alignment of multiple SAR images should be explored. Sea ice drift and alignment algorithms must be tested under various weather and drift conditions to ensure robustness. Reducing noise in Harmony data and using it to detect the MIZ, where mobile ice with high concentrations is present, will enhance detection capabilities. Practical predictability of linear kinematic features in sea ice should be estimated and improved. Developing new metrics for model calibration and validation using ice drift and deformation data is necessary. Integrating ML-based ice type products into ice service routines will aid ice analysts. Combining ice type observations with ice drift forecasts to predict ice types and ship routes will enhance navigation safety. Forecasting SAR images by integrating them with ice drift forecasts will improve operational planning. Different products should be assimilated in various regions for both operational and reanalysis purposes. The impact of assimilating different products, such as Sea Ice Concentration (SIC) and Sea Ice Drift (SoD), must be evaluated, and uncertainties characterized better. Interpreting the results of 'black box' Convolutional Neural Networks (CNNs) will transform machine learning insights into human learning. Developing a forward model for sea ice backscatter, where SAR image texture reflects the history of ice deformation, is essential for advancing sea ice physics understanding. 4.6 Targeted In-Situ Data Collection and Sharing Targeting in-situ data campaigns for satellite product validation and overlapping permanent stations with repeated satellite passes will 13 improve data quality. In-situ data collection should be tailored to address specific scientific questions, and connections between ground radar observations, drones, and SAR for upscaling should be established. Openly available multisensor API for routine overlapping in-situ + satellite sensor planning. Coordinating observations from space, drones, above sea ice, and below sea ice will enhance data integration. Improving procedures for quality control of insitu data and deploying more drifters on underrepresented sea ice types, such as FYI and fast-drifting sea ice, are critical. Increased use of georeferenced drone images for training and validation of satellite data products is recommended. Drone flights should be planned to address scientific and operational questions, and used for instantaneous sea ice drift retrieval connected with SAR observations. Drones' longdistance capabilities allow for detailed optical and infrared mapping below cloud cover and simultaneous with SAR acquisitions. By addressing these recommendations, significant advancements can be made in the field of SAR remote sensing of sea ice, enhancing the accuracy and reliability of sea ice monitoring and forecasting. REFERENCES Barber D., E. LeDrew, D. Flett, M. Shokr, and J. Falkingham, “Seasonal and diurnal variations in SAR signatures of landfast sea ice,” IEEE transactions on geoscience and remote sensing, vol. 30, no. 3, pp. 638–642, 1992. Casey J. A., S. E. Howell, A. Tivy, and C. Haas, “Separability of sea ice types from wide swath Cand L-band synthetic aperture radar imagery acquired during the melt season,” Remote Sensing of Environment, vol. 174, pp. 314 – 328, 2016. Demchev, D., L. E. B. Eriksson, A. Hildeman and W. Dierking: "Alignment of Multifrequency SAR Images Acquired Over Sea Ice Using Drift Compensation,", IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 7393-7402, 2023, doi: 10.1109/JSTARS.2023.3302576. Doulgeris, A.P.: An Automatic U-Distribution and Markov Random Field Segmentation Algorithm for PolSAR Images, IEEE Transactions on Geoscience and Remote Sensing, vol. 53, no. 4, pp. 1819–1827, 2015. Frost, A., Wiehle, S., Singha, S., & Krause, D. (2018, July). Sea Ice Motion Tracking from Near Real Time Sar Data Acquired During Antarctic Circumnavigation Expedition. In IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium (pp. 2338-2341). IEEE. Horn, B. K. P. and Schunck, B. G., Determining optical flow, Artif. Intell., 17, 185–203, 1981. Johansson A. M., C. Brekke, G. Spreen, and J. A. King, “X-, C-, and L-band SAR signatures of newly formed sea ice in Arctic leads during winter and spring,” Remote Sensing of Environment, vol. 204, pp. 162 – 180, 2018. Karvonen, J., Baltic Sea Ice Concentration Estimation Using SENTINEL-1 SAR and AMSR2 Microwave Radiometer Data, in IEEE Transactions on Geoscience and Remote Sensing, vol. 55, no. 5, pp. 28712883, May 2017, doi: 10.1109/TGRS.2017.2655567. Karvonen, J., Baltic Sea Ice Concentration Estimation From C-Band Dual-Polarized SAR Imagery by Image Segmentation and Convolutional Neural Networks, in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-11, 2022, Art no. 4301411, doi: 10.1109/TGRS.2021.3097885. Kleinherenbrink, M.: Harmony sea-ice dynamics code and Notebooks, Zenodo, https://doi.org/10.5281/zenodo.5055737, 2021a. Kleinherenbrink, M., Korosov, A., Newman, T., Theodosiou, A., Komarov, A. S., Li, Y., Mulder, G., Rampal, P., Stroeve, J., and Lopez-Dekker, P.: Estimating instantaneous sea-ice dynamics from space using the bistatic radar measurements of Earth Explorer 10 candidate Harmony, The Cryosphere, 15, 3101–3118, https://doi.org/10.5194/tc-153101-2021, 2021b. Kleinherenbrink, M., Korosov, A., Newman, T., Theodosiou, A., Komarov, A. S., Li, Y., Mulder, G., Rampal, P., Stroeve, J., and Lopez-Dekker, P.: Estimating instantaneous sea-ice dynamics from space using the bi- 14 static radar measurements of Earth Explorer 10 candidate Harmony, The Cryosphere, 15, 3101–3118, https://doi.org/10.5194/tc-153101-2021, 2021b. König C., T. König, S. Singha, A. Frost, and S. Jacobsen, “Combined ¨ Use of Space Borne Optical and SAR Data to Improve Knowledge about Sea Ice for Shipping,” Remote Sensing, vol. 13, no. 23, 2021. Kortum, K., Singha, S., & Spreen, G. (2022). Robust Multiseasonal Ice Classification From High-Resolution X-Band SAR. IEEE Transactions on Geoscience and Remote Sensing, 60, 1-12. Kræmer T., H. Johnsen, C. Brekke and G. Engen, "Comparing SAR-Based Short TimeLag Cross Correlation and Doppler-Derived Sea Ice Drift Velocities," in IEEE Transactions on Geoscience and Remote Sensing, vol. 56, no. 4, pp. 1898-1908, April 2018, doi: 10.1109/TGRS.2017.2769222. keywords: {Doppler effect;Satellites;Sea ice;Synthetic aperture radar;Antennas;Global Positioning System;Doppler measurement;motion estimation;sea ice;synthetic aperture radar}. Krumpen, Thomas, et al. "Arctic warming interrupts the Transpolar Drift and affects long-range transport of sea ice and ice-rafted matter." Scientific Reports 9.1 (2019): 5459. Kwok, R.: The RADARSAT Geophysical Processor System, in: Analysis of SAR Data of the Polar Oceans, Springer Berlin Heidelberg, 235–257, https://doi.org/10.1007/978-3-642-602825_11, 1998. a, b Lohse, J and Doulgeris, A.P. and Dierking, W.: Mapping sea-ice types from Sentinel-1 considering the surface-type dependent effect of incidence angle, Annals of Glaciology, vol. 61, no. 83, pp. 260–270, https://doi.org/10.1017/aog.2020.45, 2020. Lucas, B. and Kanade, T., An iterative image registration technique with an application to stereo vision, in: Proceedings of the 7th International Joint Conference on Artificial Intelligence (IJCAI), 24–28 August 1981, Vancouver, Canada, vol. 2, 674–679, 1981. Mahmud M., V. Nandan, S. Howell, T. Geldsetzer, and J. Yackel, “Seasonal evolution of l-band SARbackscatter over landfast arctic sea ice,” Remote Sensing of Environment, vol. 251, p. 112049, 2020. Murashkin D., and A. Frost, “Arctic Sea Ice Mapping Using Sentinel-1 SAR Scenes with a Convolutional Neural Network,” IEEE International Geoscience and Remote Sensing Symposium, pp. 5660-5663. 2021. Olason, E., Boutin, G., Korosov, A., Rampal, P., Williams, T., Kimmritz, M., et al. (2022). A new brittle rheology and numerical framework for large-scale sea-ice models. Journal of Advances in Modeling Earth Systems, 14, e2021MS002685. https://doi.org/10.1029/2021MS002685. Panchi N., E. Kim and A. Bhattacharyya, "Supplementing Remote Sensing of Ice: Deep Learning-Based Image Segmentation System for Automatic Detection and Localization of Sea-ice Formations From Close-Range Optical Images," IEEE Sensors Journal, vol. 21, no. 16, pp. 18004-18019, 2021. Perona P., J. Malik, Scale space and edge detection using anisotropic diffusion, Proc. IEEE Comp. Soc. Workshop on Computer Vision (Miami Beach, Nov. 30 – Dec. 2, 1987), IEEE Computer Society Press, Washington, 16–22, 1987. Rublee, E., V. Rabaud, K. Konolige and G. Bradski, ORB: An efficient alternative to SIFT or SURF, 2011 International Conference on Computer Vision, Barcelona, Spain, 2011, pp. 2564-2571, doi: 10.1109/ICCV.2011.6126544. Toyota T., J. Ishiyama, and N. Kimura, “Measuring deformed sea ice in seasonal ice zones using l-band SAR images,” IEEE Trans. Geosci. Remote Sens., vol. 59, no. 11, pp. 9361–9381, 2020. Wang R., W. Zhu, X. Zhang, Y. Zhang and J. Zhu, "Comparison of Doppler-Derived Sea Ice Radial Surface Velocity Measurement Methods From Sentinel-1A IW Data," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 16, pp. 2178-2191, 2023. Yackel J., and D. Barber, “Melt ponds on sea ice in the Canadian Archipelago: 2. On the use of RADARSAT-1 synthetic aperture radar for geophysical inversion,” Journal of [Document text truncated for crawler view.]